One of the things I often wonder is "what will be the minimally viable LLM" that can work from just enough information that if it googles the rest it can provide reasonable answers? I'm surprised something like Encyclopedia Britanica hasn't yet (afaik) tried to capitalize on AI by selling their data to LLMs and validating outputs for LLM companies, it would make a night and day difference in some areas I would think. Wikipedia is nice, but there's so much room for human error and bias there.
Also this is the direction the small LLMs are moving in already. They are too small for general knowledge, but getting quite good at tool use (incl. Googling).
Now we just need them to be very strict about what they know and don't know! (I think this is still an open problem, even with big ones.)
It's not so much a "minimally viable LLM" but rather an LLM that knows natural language well but knows nothing else. Like me - as an engineer who knows how to troubleshoot in general but doesn't know about a specific device like my furnace (recent example).
And I don't think that LLM could just Google or check Wikipedia.
But I do agree that this architecture makes a lot of sense. I assume it will become the norm to use such edge LLMs.
I asked this question a while back (the "only train w/ wikipedia LLM") and got pointed to the general-purpose "compression benchmarks" page: `https://www.mattmahoney.net/dc/text.html`
While I understand some of the fundamental thoughts behind that comparison, it's slightly wonky... I'm not asking "compress wikipedia really well", but instead "can a 'model' reason its way through wikipedia" (and what does that reasoning look like?).
Theoretically with wikipedia-multi-lang you should be able to reasonably nail machine-translation, but if everyone is starting with "only wikipedia" then how well can they keep up with the wild-web-trained models on similar bar chart per task performance?
If your particular training technique (using only wikipedia) can go from 60% of SOTA to 80% of SOTA on "Explain why 6-degrees of Kevin Bacon is relevant for tensor operations" (which is interesting to plug into Google's AI => Dive Deeper...), then that's a clue that it's not just throwing piles of data at the problem, but instead getting closer to extracting the deeper meaning (and/or reasoning!) that the data enables.
Correct! I know RAG is a thing, but I wish we could have "DLCs" for LLMs like image generation has LoRa's which are cheaper to train for than retraining the entire model, and provide more output like what you want. I would love to pop in the CS "LoRa or DLC" and ask it about functional programming in Elixir, or whatever.
Maybe not crawl the web, but hit a service with pre-hosted, precurated content it can digest (and cache) that doesn't necessarily change often enough. You aren't using it for the latest news necessarily, but programming is mostly static knowledge a a good example.
Your worry about Wikipedia is that there is "much room for human error and bias", yet earlier you seem to imply that a LLM that has access to the www somehow would have less human error and bias? Personally, I'd see it the other way around.
Isn’t that sort of what a RAG is? You’d need an LLM “smart” enough to turn natural-user prompts into searches, then some kind of search, then an LLM “smart” though to summarize the results.
Yeah, I think RAG is the idea that will lead us there, though its a little complicated, because for some subjects, say Computer Science, you need a little more than just "This is Hello World in Go" you might need to understand not just Go syntax on the fly, but more CS nuances that are not covered in one single simple document. The idea being having a model that runs fully locally on a phone or laptop with minimal resources. On the other hand, I can also see smaller models talking to larger models that are cheaper to run in the cloud. I am wondering if this is the approach Apple might take with Siri, specifically in order to retain user privacy as much as possible.
I remember reading tht hallucination is still a problem even with perfect context. You build a theoretical perfect RAG, give the LLM the exact correct information, and it will still make mistakes surprisingly often.
Unfortunately reasoning ability depends on (or is enabled by) information intake during training. A model will know better what to search for and how to interpret it if the information was part of the training. So there is a trade off. Still I think the question is a practical one. Perhaps there are ideas to focus training on a) reasoning / conceptual modeling and b) reliance on external memory (search etc.) rather than internal memorization.
How? They can validate thousands if not millions of queries but nothing prevent the millions-th-and-one from being a hallucination. People who would then pay extra for a "Encyclopedia Britanica validated LLM" would then, rightfully so IMHO, complain that "it" suggested them to cook with a dangerous mushroom.
> I often wonder is "what will be the minimally viable LLM" that can work from just enough information that if it googles the rest it can provide reasonable answers?
It depends what that word "reasonable" means for your specific use-case ;)
Wikipedia has proven to be as accurate as encyclopedias for decades now. Also, I'm betting AI companies have illegally trained their models on the Encyclopedia Britanica's data by now.
It's perfectly legal to train a human on copyrighted work and I think, depending on the country, it's not settled that training ai on the same data is illegal.
I think the idea is to train a small, minimal LLM thinking model that can run on edge devices, but that has very little knowledge embedded in its weights, and so performs a sort of RAG to Encylopedia Britannica to ground answers to user queries.
https://arxiv.org/pdf/2310.11453
The original paper [fig 1, bottom-right] seems to say it needs about 4-5x the parameters of a fp16 model. You can build it and run some models, but the selection is limited because it has to be trained from scratch. I imagine inference speed is faster compared with modern PTQ (4- and 8-bit quants) though.
So Microsoft is actually using 2 bits instead of 1.58. In this case they could represent -1, 0, 1, 2. As inhibitory synapses account for 20%-30%, this could map well to how biological brains are structured.
In the human brain most synapses are indeed excitatory, while a minority is inhibitory.
No concise HN comment will give you a complete picture of whats currently known about the human brain, so a platitude necessarily follows:
We call the nearly touching interfaces between neurons synapses, small packets / droplets of neurotransmitter are sent across this interface from the source to the target neuron. Such signals can be excitatory (promote the probability of excitation of the target firing soon) or inhibitory (inhibits the probability of the target firing soon). There are 2 types of sensitive areas on your average neuron: the dendrites (long branching tentacles, that receive excitatory signals) and the cell body where all the signals are accumulated to a local instantaneous "sum" is also sensitive to synaptic activation, but the synapses on the cell body are inhibitory, when sufficiently inhibited the neuron will refuse to fire its axons, so the inhibitory synapses on the cell body can gate the cumulative signal and prevent it from triggering this neuron temporarily. If the neuron does fire, this propagates along the axons (another type of branching tentacles, which lead to yet other neurons, sometimes touching them excitatorily at their dendrite, sometimes touching a neuron inhibitorily at their cell body.
It is really truly incredible that this mess of microscopic meat plumbing encodes everything we see, think, and do. Terrifying and amazing all at once.
I did not realize all the dendritic synapses were excitatory, I always thought it depended on the specific neurotransmitters released. Thanks, this is cool. I am curious what will happen when we build LLMs that have the equivalent of chemical diffusions between synaptic release areas as well as the temporality of spiking neural nets.
It's good to see this getting some continued development. I looked into it last year[1] and I thought it showed a lot of promise so I've been very disappointed that I never saw a newer model.
I think this approach is not so interesting because it's just quantization of a full precision model. So it speeds up inference (at a quality penalty) but not training. It would be more interesting to train an actually binary model directly, without any floating point multiplication, like in this paper: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7...
The energy numbers are the real story here, 70-82% reduction on CPU inference. If 1-bit models ever get good enough, running them on commodity hardware with no GPU budget changes who can deploy LLMs. That's more interesting than the speed benchmarks imo.
> bitnet.cpp is the official inference framework for 1-bit LLMs (e.g., BitNet b1.58). It offers a suite of optimized kernels, that support fast and lossless inference of 1.58-bit models on CPU and GPU (NPU support will coming next).
Log Base 2 of 3 = ~1.5849625, so that's the limit to how well you can pack three-state values into bits of data.
For something more practical, you can pack five three-state values within a byte because 3^5 = 243, which is smaller than 256. To unpack, you divide and modulo by 3 five separate times. This encodes data in bytes at 1.6 bits per symbol.
But the packing of 5 symbols into a byte was not done here. Instead, they packed 4 symbols into a byte to reduce computational complexity (no unpacking needed)
Yeah, "1.58 bit" is 1 trit with three states, since log2(3)≈1.58.
So it's not a inference framework for 1-bit models (two states per parameter) but for 1.58 bit models (three states per parameter). Annoying that they try to mix up the two.
It might interest you to know that one or two months ago, I had Claude port BitNet to WebGPU from the reference implementation, so that it runs right in your browser as a local model. After some debugging, the port seemed to work, but the model didn't function as well as the reference implementation so I'll have to work on it for a while. You can see a debugging session livestreamed here[1]. The released model file was about a gigabyte, it fits in most people's GPU's. We were also able to successfully fine-tune it right in the browser.
There's a lot that you can do when the model size is that small, yet still powerful.
Our next step is that we want to put up a content distribution network for it where people can also share their diffs for their own fine-tuned model. I'll post the project if we finish all the parts.
I think they used a dummy model or else they would have linked to it. Just google '1-bit 100b model' and you'll only see references to this project without any download links.
The title is misleading — there's no trained 100B model, just an inference framework that claims to handle one. But the engineering is worth paying attention to.
I run quantized 70B models locally (M2 Max 96GB, llama.cpp + LiteLLM), and memory bandwidth is always the bottleneck. The 1.58-bit approach is interesting because ternary weights turn matmuls into additions — a fundamentally different compute profile on commodity CPUs. If 5-7 tok/s on a single CPU for 100B-class models is reproducible, that's a real milestone for on-device inference.
Framework is ready. Now we need someone to actually train the model.
> Framework is ready. Now we need someone to actually train the model.
If Microslop aren't gonna train the model themselves to prove their own thesis, why would others? They've had 2 years (I think?) to prove BitNet in at least some way, are you really saying they haven't tried so far?
Personally that makes it slightly worrisome to just take what they say at face value, why wouldn't they train and publish a model themselves if this actually led to worthwhile results?
Because this is Microsoft, experimenting and failing is not encouraged, taking less risky bets and getting promoted is. Also no customer asked them to have 1-bit model, hence PM didn't prioritize it.
But it doesn't mean, idea is worthless.
You could have said same about Transformers, Google released it, but didn't move forward, turns out it was a great idea.
> You could have said same about Transformers, Google released it, but didn't move forward,
I don't think you can, Google looked at the research results, and continued researching Transformers and related technologies, because they saw the value for it particularly in translations. It's part of the original paper, what direction to take, give it a read, it's relatively approachable for being a machine learning paper :)
Sure, it took OpenAI to make it into an "assistant" that answered questions, but it's not like Google was completely sleeping on the Transformer, they just had other research directions to go into first.
> But it doesn't mean, idea is worthless.
I agree, they aren't, hope that wasn't what my message read as :) But, ideas that don't actually pan out in reality are slightly less useful than ideas that do pan out once put to practice. Root commentator seems to try to say "This is a great idea, it's all ready, only missing piece is for someone to do the training and it'll pan out!" which I'm a bit skeptical about, since it's been two years since they introduced the idea.
What OpenAI did was train increasingly large transformer model instances. which was sensible because transformers allowed for a scaling up of training compared to earlier models. The resulting instances (GPT) showed good understanding of natural language syntax and generation of mostly sensible text (which was unprecedented at the time) so they made ChatGPT by adding new stages of supervised fine tuning and RLHF to their pretrained text-prediction models.
There were plenty of models the size of gpt3 in industry.
The core insight necessary for chatgpt was not scaling (that was already widely accepted): the insight was that instead of finetuning for each individual task, you can finetune once for the meta-task of instruction following, which brings a problem specification directly into the data stream.
Google had been working on a big LLM but they wanted to resolve all the safety concerns before releasing it. It was only when OpenAI went "YOLO! Check this out!" that Google then internally said, "Damn the safety concerns, full speed ahead!" and now we find ourselves in this breakneck race in which all safety concerns have been sidelined.
Scaling seemed like the important idea that everyone was chasing. OpenAI used to be a lot more safety minded because it was in their non profit charter, now they’ve gone for-profit and weaponized their tech for the USA military. Pretty wild turnaround. Saying OpenAI was cavalier with safety in the early days is inaccurate. It was a skill issue. Remember Bard? Google was slow.
On the one hand, not publishing any new models for an architecture in almost a year seems like forever given how things are moving right now. On the other hand I don't think that's very conclusive on whether they've given up on it or have other higher priority research directions to go into first either
> You could have said same about Transformers, Google released it, but didn't move forward, turns out it was a great idea
Google released transforms as research because they invented it while improving Google Translate. They had been running it for customers for years.
Beyond that, they had publicly-used transformer based LMs ("mums") integrated into search before GPT-3 (pre-chat mode) was even trained. They were shipping transformer models generating text for years before the ChatGPT moment. Literally available on the Google SERP page is probably the widest deployment technology can have today.
Transformers are also used widely in ASR technologies, like Google Assistant, which of course was available to hundreds of millions of users.
Finally, they had a private-to-employees experimental LLMs available, as well as various research initatives released (meena, LaMDA, PaLM, BERT, etc) and other experiments, they just didn't productize everything (but see earlier points). They even experimented with scaling (see "Chinchilla scaling laws").
The most benign answer would be that they don’t want to further support an emerging competitor to OpenAI, which they have significant business ties to. I think the more likely answer which you hinted at is that the utility of the model falls apart as scale increases. They see the approach as a dead end so they are throwing the scraps out to the stray dogs.
Not to mention Microsoft's investments in Nvidia and other GPU-adjacent/dependent companies!
A successful ternary model would basically erase all that value overnight. In fact, the entire stock market could crash!
Think about it: This is Microsoft we're talking about! They're a convicted monopolist that has a history of manipulating the market for IT goods and services. I wouldn't put it past them to refuse to invest in training a ternary
model or going so far as to buy up ternary startups just to shut them down.
Want to make some easy money: Start a business training a ternary model and make an offer to Microsoft. I bet they'll buy you out for at least a few million even if you don't have a product yet!
Rest assured, all the big players (openai, google, deepseek etc) have run countless experiments with 4,3,2,1.58,1 bits, and various sparse factors and shapes. This barrel has been scraped to the bottom
I have doubts about this. Perhaps the closed models have, but I wouldn't be so sure for the open ones.
GLM 5, for example, is running 16-bit weights natively. This makes their 755B model 1.5TB in size. It also makes their 40B active parameters ~80GB each.
Compare this to Kimi K2.5. 1T model, but it's 4-bit weights (int4), which makes the model ~560 GB. Their 32B active parameters are ~16 GB.
Sure, GLM 5 is the stronger model, but is that price worth paying with 2-3x longer generation times? What about 2-3x more memory required?
I think this barrel's bottom really hasn't been scraped.
The title being misleading is important as well, because this has landed on the front page, and the only thing that would be the only notable part of this submission.
The "new" on huggingface banner has weights that were uploaded 11 months ago, and it's 2B params. Work on this in the repo is 2 years old.
The amount of publicity compared to the anemic delivery for BitNet is impressive.
I've also always though that it's an interesting opportunity for custom hardware. Two bit addition is incredibly cheap in hardware, especially compared to anything involving floating point. You could make huge vector instructions on the cheap, then connect it to the fastest memory you can buy, and you have a capable inference chip.
You'd still need full GPUs for training, but for inference the hardware would be orders of magnitude simpler than what Nvidia is making
These are trits, which provide their own efficiencies.
Interestingly, a trit x float multiplier is cheaper than a trit x integer multiplier in hardware if you're willing to ignore things like NaNs.
0 and 1 are trivial, just a mux for identity and zero. But because floats are sign-magnitude, multiply by -1 is just an inverter for the sign bit, where as for integers you need a bitwise inverter and full incrermenter.
There's lots of literature on quantizing weights (including trits and binary) going back 15+ years. Nothing to hand right now but it's all on arxiv.
The relevant trit arithmetic should be on display in the linked repo (I haven't checked). Or try working it out for the uncompressed 2 bit form with a pen and paper. It's quite trivial. Try starting with a couple bitfields (inputs and weights), a couple masks, and see if you can figure it out without any help.
You only need GPUs if you assume the training is gradient descent. GAs or anything else that can handle nonlinearities would be fine, and possibly fast enough to be interesting.
Text is misleading too. 5-7 tok/sec is not reading speed, it's a tad slower. For me, at least, and I am an experienced reader, not especially schooled in quick-reading though.
I happened to "live" on 7.0-7.5 tok/sec output speed for a while, and it is an annoying experience. It is the equivalent of walking behind someone slightly slower on a footwalk. I dealt with this by deliberately looking away for a minute until output was "buffered" and only then started reading.
For any local setup I'd try to reach for 10 tok/sec. Sacrifice some kv cache and shove a few more layers on your GPU, it's worth it.
> a fundamentally different compute profile on commodity CPU
In what way? On modern processors, a Fused Multiply-Add (FMA) instruction generally has the exact same execution throughput as a basic addition instruction
You drop the memory throughput requirements because of the packed representation of bits so an FMA can become the bottleneck, and you bypass the problem of needing to upscale the bits to whatever FP the FMA instruction needs.
typically for 1-bit matmul, you can get away with xors and pop_counts which should have a better throughput profile than FMA when taking into account the SIMD nature of the inputs/outputs.
It can probably be made more efficient by taking a column-first format.
Since we are in CPU land, we mostly deal with dot products that match the cache size, I don't assume we have a tiled matmul instruction which is unlikely to support this weird 1-bit format.
Haven't looked closely, but on modern x86 CPUs it might be possible to do much better with the gf2affineqb instructions, which let us do 8x8 bit matrix multiplications efficiently. Not sure how you'd handle the 2-bit part, of course.
The win is in how many weights you process per instruction and how much data you load.
So it's not that individual ops are faster — it's that the packed representation lets each instruction do more useful work, and you're moving far less data from memory to do it.
I'm hoping that today's complaints are tomorrow's innovations. Back when 1Mb hard drive was $100,000, or when Gates said 640kb is enough.
Perhaps some 'in the (chip) industry' can comment on what RAM manufacturers are doing at the moment - better, faster, larger? Or is there not much headroom left and it's down to MOBO manufacturers, and volume?
Chip speed has increased faster than memory speed for a long time now, leaving DRAM behind. GDDR was good for awhile but is no longer sufficient. HBM is what's used now.
The last logical step of this process would be figuring out how to mix the CPU transistors with the RAM capacitors on the same chip as opposed to merely stacking separate chips on the same package.
A related stopgap is the AI startup (forget which) making accelerators on giant chips full of SRAM. Not a cost effective approach outside of ML.
We have faster memory, it's just all used in data center cards you can't buy (and can't afford to buy).
AMD actually used HBM2 memory in their Radeon VII card back in 2019 (!!) for $700. It had 16 GB of HBM2 memory with 1 TB/s throughput.
The RTX 5080 in conversion l comparison also has 16 GB of VRAM, but was released in 2025 and has 960 GB/s throughput. The RTX 5090 does have an edge at 1.8 TB/s bandwidth and 32 GB of VRAM but it also costs several times more. Imagine if GPUs had gone down the path of the Radeon VII.
That being said, the data center cards from both are monstrous.
The Nvidia B200 has 180 GB of VRAM (2x90GB) offering 8.2 TB/s bandwidth (4.1 TB/s x2) released in 2024. It just costs as much as a car, but that doesn't matter, because afaik you can't even buy them individually. I think you need to buy a server system from Nvidia or Dell that will come with like 8 of these and cost you like $600k.
AMD has the Mi series. Eg AMD MI325x. 288 GB of VRAM doing 10 TB/s bandwidth and released in 2024. Same story as Nvidia: buy from an OEM that will sell you a full system with 8x of these (and if you do get your hands on one of these you need a special motherboard for them since they don't do PCIe). Supposedly a lot cheaper than Nvidia, but still probably $250k.
These are not even the latest and greatest for either company. The B300 and Mi355x are even better.
It's a shame about the socket for the Mi series GPUs (and the Nvidia ones too). The Mi200 and Mi250x would be pretty cool to get second-hand. They are 64 GB and 128GB VRAM GPUs, but since they use OAP socket you need the special motherboard to run them. They're from 2021, so in a few years time they will likely be replaced, but as a regular joe you likely can't use them.
The systems exist, you just can't have them, but you can rent them in the cloud at about $2-4 per hour per GPU.
For larger contexts, the bottleneck is probably token prefill instead of memory bandwidth. Supposedly prefill is faster on the M5+ GPUs, but still a big hurdle for pre-M5 chips.
Yes. I had to read it over twice, it does strike me as odd that there wasn't a base model to work with.
But it seems the biggest model available is 10B? Somewhat unusual and does make me wonder just how challenging it will be to train any model in the 100B order of magnitude.
Approximately as challenging as training a regular 100B model from scratch. Maybe a bit more challenging because there's less experience with it
The key insight of the BitNet paper was that using their custom BitLinear layer instead of normal Linear layers (as well as some more training and architecture changes) lead to much, much better results than quantizing an existing model down to 1.58 bits. So you end up making a full training run in bf16 precision using the specially adapted model architecture
What's unusual about it? It seems pretty standard to train small models to validate an approach, and then show that training scales with model size to 8B to 14B parameter models, which is what they did.
There are 1 bit average GGUFs of large models, not perfect quality but they will hold a conversation. These days, there is also quantized finetuning to heal the damage.
That issue appears to be the one that's wrong. From the technical report
> We evaluated bitnet.cpp in terms of both inference speed and energy cost. Comprehensive tests were conducted on models with various parameter sizes, ranging from 125M to 100B. specific configurations for each model are detailed in the Appendix A.
Thanks for pointing that out. I'll ask the issue creator if they've considered that. Would be nice if the maintainer would handle that (sigh) and link to the actual models used for testing (double sigh).
From what I gather, there are no models, this is a framework for running 1bit models, but none have been trained. They are mainly demonstrating the possibility.
I also don't expect those with poor MCPs to have any better CLIs or APIs, most of the big companies we want them for are not investing in DX/AX. I suspect i.e. that Intuit, if they had great APIs et al, would see it as a threat to their business.
Boy would I love to give my agent access to my Quickbooks. They pushed out an incomplete MCP and haven't touched it since.
Hmm, the user joined in 2019 but had no submissions or comments until just 40 minutes ago (at least judging by the lack of a second page?) and all the comments are on AI related submissions. Benefit of doubt is it'd have to be a very dedicated lurker or dormant account they remembered they had.
Edit: oh, just recalled dang restricted Show HNs the other day to only non-new users (possibly with some other thresholds). I wonder if word got out and some are filling accounts with activity.
There has been a shift to the Ai accounts, they use Show HN less now. This started before dang's comment, I assume because they saw the earlier posts about the increase in quantity / decrease in quality.
I suspect that they are trying to fake engagement prior to making their first "show" post as well.
Fair enough — I've been lurking since 2019 and picked a bad day to start commenting on everything at once. Not a bot, just overeager. I'll pace myself.
Your account posted dense, opinionated and structured paragraphs mere minutes apart—sometimes the same minute—for multiple story submissions. Even with my own sometimes lengthy replies this would be infeasible to both instantly have structured opinions and type them out in time. Two of your posts were posted the same minute, with a combined word count of 146.
It feels like it'd take someone superhuman to come across different stories, have such opinions and type and submit both of these in that timeframe or queuing up comments to post rapid-fire.
Conspicuously too, as another pointed out, is every single comment of yours uses an em dash, which despite occasionally using myself (hey look they're in this reply) is not in every single comment. Idk, if I was being seriously accused of botting I'd put more reasoning into my response about it.
Lol. I know at least a few high karma account who post at the same frequency but they post about anti-AI and anti-tech topics instead on the big social media tech where anti-tech opinions dominate. I guess this exempts them from scrutiny? I love these witch hunts.
They post 146 words per minute across multiple different submissions with similarly structured posts? I know there are users who post frequently in other communities I'm familiar with but not in that kind of timeframe with such paragraph density or structure.
Yes. Any of the really heated political threads on this site are full of posters like this. I don't want to dox anyone, but since at least 2019-ish I've found posters that spend hours a day posting huge amounts of content in large bursts on this site.
COVID was ridiculous as I presume a lot of anxious people were stuck at home able to do nothing but post.
Does this account for them reading the article though? There are pre-existing opinions that could be easy to rapidly post based on topical discussions but here it also has to be considered the time to parse technical submissions.
They read this article, called out a specific discrepancy then commented on a paper on Arxiv in a 70-odd word post then the same minute another 70-odd word post on a different technical post. Maybe like you suggest they're just wired differently.
Yeah, there are plenty of very online people who are great at just posting things and doing the minimum amount of diligence to contribute their thought. Like I said, just hang around in any of the high energy threads on this site and you'll see a lot of it.
It's scary, without the em dashes, and the rapid fire commenting of the account - who would ever realize this is a bot? Two easy to fix things, and after that it'd be very difficult to tell that this is a bot.
It's not a question of if there are other bots out there, but only what % of comments on HN right now and elsewhere are bot generated. That number is only going to increase if nothing is done.
Looks like gradual disempowerment is already happening - the minority of humans who are capable of spotting AI content are losing the struggle for attention on all major social networks
Funny enough I now involuntarily take RTFA as a slight slop signal, because all these accounts dutifully read the article before commenting, unlike most HNers who often respond to headlines.
I’ve been rounded up for things I wrote two decades ago because of my em dashes lol. The pitchfork mentality gives me little hope for how things are going to go once we have hive mind AGI robots pervasive in society.
If I was operating a bot farm, at this point I would probably add some bots that go around and accuse legit human users (or just random users) of being bots.
Created confusion and frustration will make it much harder to separate signal from the noise for most people.
I once spent some time learning the proper usage of em-dashes, en-dashes, and hyphens, and tried to be conscientious about using them properly in my writing. Little did I know it would be wasted effort in the LLM era, when competent writing actually became a negative.
Not only are we losing the ability to communicate clearly without the assistance of computers, those who can are being punished for it.
Not all of them do: https://news.ycombinator.com/item?id=47335156 There are evidently lots of people experimenting with different botting setups. Some do better at blending in than others.
Interesting - the account you mention, and the GP, are both doing replies that are themselves all about the same length, and also the same length between the two accounts. I get what you mean.
I would love to understand the thought process behind this. I'm sure it's a fun experiment, to see if it's possible and so on... but what tangible benefit could there be to burning tokens to spam comments on every post?
The output from this model is horrible! It's GPT-2 level babble and repeats entire paragraphs verbatim. It also reuses the same fake citation `(Jenkins, 2010)` over and over again. From the start of their video (which scrolls by fast enough that you don't see the slop clearly...)
```
Ecosystem Services and their impact on the Ecosystem
Ecosystem services refer to the services provided by ecosystems to the human society. These services include water, air, energy, nutrients, and soil (Jenkins, 2010). For instance, water is the most important service provided by an ecosystem and it helps in the conservation of water, irrigation and sanitation (Jenkins, 2010). On the other hand, air provides the oxygen needed for life.
The water cycle is a significant ecosystem service because it involves the cycling of water among the different parts of an ecosystem. It also involves the movement of water through the atmosphere, from one place to another. It is also the process of evaporation and condensation of water from the atmosphere. It also involves the movement of water from the air to the soil and water into the oceans.
The water cycle is a significant ecosystem service because it involves the cycling of water among the different parts of an ecosystem. It also involves the movement of water through the atmosphere, from one place to another. It is also the process of evaporation and condensation of water from the atmosphere. It also involves the movement of water from the air to the soil and water into the oceans.
```
Thanks for the link, the GSM8K result actually leads the pack in that table, but math is indeed underwhelming. Qwen 2.5 is in the lead, but bitnet isn't far behind and it takes 1/6th as much memory during inference, and was trained on less than 1/4 the number of tokens. Pretty cool.
I think the README [1] for the new CPU feature is of more interest, showing linear speedups with number of threads. Up to 73 tokens/sec with 8 threads (64 toks/s for their recommended Q6 quant):
I wonder when we begin to see the dividends of all the NPU PCs come into play. AMD have been doing some good work with their NPU/iGPU hybrid inference kernels. If these larger models could be scaled down to run on NPUs, you'd see much better power advantages, compared to running them on the CPU.
> I wonder when we begin to see the dividends of all the NPU PCs come into play.
A few months ago I used Whisper from OpenAI, an automatic speech recognition system released in 2002, on my modern 20-core Intel CPU to convert audio from a video file to text. It worked fine. Took a while and the machine got hot and the fans kicked in. I then found the Intel's optimized version of whisper that used NPU. It required a lot more steps to get working, but in the end it did work and was about 6x faster. And the machine remained cool and silent in the process. Since then I have become a fan of the NPUs. They are not NVIDIA GeForce RTX 5090, but they are significantly better than a modern CPU.
You can already run some models on the NPUs in the Rockchip RK3588 SBCs which are pretty abundant.
A claude 4.6 they are most certainly not, but if you get through the janky AF software ecosystem they can run small LLMs reasonably well with basically zero CPU/GPU usage
I'm curious if 1-bit params can be compared to 4- or 8-bit params. I imagine that 100B is equivalent to something like a 30B model? I guess only evals can say. Still, being able to run a 30B model at good speed on a CPU would be amazing.
At some point you hit information limits. With conventional quantisation you see marked capability fall-off below q5. All else being equal you'd expect an N-parameter 5-bit quant to be roughly comparable to a 3N-parameter ternary, if they are trained to the same level, just in terms of the amount of information they can possibly hold. So yes, 100B ternary would be within the ballpark of a 30B q5 conventional model, with a lot of hand-waving and sufficiently-smart-training
I assume that theoretically, 1-bit models could be most efficient because modern models switched from 32 bit to 16 bit to 8 bit per parameter (without quantization).
It's not clear where the efficiency frontier actually is. We're good at measuring size, we're good at measuring FLOPS, we're really not very good at measuring capability. Because of that, we don't really know yet whether we can do meaningfully better at 1 bit per parameter than we currently get out of quantising down to that size. Probably, is the answer, but it's going to be a while before anyone working at 1 bit per param has sunk as many FLOPS into it as the frontier labs have at higher bit counts.
The thing with efficiency is that it is relative to both inference and training compute. If you do quantization, you need a more powerful higher precision model to quantize from, which doesn't exist if you want to create a frontier model. In this case the question is only whether you get better inference and/or training performance from training e.g. a native 1 bit model.
Currently the optimal training precision seems to be 8 bit (at least used by DeepSeek and some other open weight companies). But this might change with different training methods optimized for 1-bit training, like from this paper I linked before: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7...
I don't see the news here ... there's https://huggingface.co/collections/microsoft/bitnet which is last updated 12/2025 ... am I just paying more attention here or is there something actually new about this?
Also as far as I know, this is more of a research curiosity - BitNet really doesn't perform that well on evals.
I think Qwen3.5 2B is the best you can get in the ~1GB class.
The project is an inference framework which should support 100B parameter model at 5-7tok/s on CPU. No one has quantized a 100B parameter model to 1 trit, but this existing is an incentive for someone to do so.
I had the same question, after some debates with Chatgpt, it's not the "quantize" for post-training we often witness these days, you have to use 1 trit in the beginning since pre-train.
If they had a big result like, native 1.58-bit quality clearly matches top peers, they would be saying that prominently in the repo.
The engineering/optimization work is nice, but this is not what people have been waiting for, as much as, can’t the Bitnet idea that seemed promise really deliver in a competitive way.
If this stuff was so revolutionary, don't you guys think Qwen/DeepSeek would have snapped it up already? Both those teams are highly innovative, picking up and inventing new techniques all the time. Hell, Deepseek-v3 was one of the first to do large scale fp8 training.
Do LLMs have a way to look at or consider dependent variables?
Seems like that could end up as a situation where a fractional number of bits or bytes per parameter might make sense. Particularly with adverbs and adjectives, negators.
Misleading title but this is pretty exciting. Interesting how this is based on llama cpp. Its nice to see some momentum since they released the paper in 2023
They have a demo video in the readme. I think they are trying to convey that BitNet is fast, which it is. But it is worth taking a moment to pause and actually see what the thing is doing so quickly.
It seems to keep repeating that the water cycle is the main source of energy for all living things on the planet and then citing Jenkins 2010. There are also a ton of sentence beginning with “It also…”
I don’t even think it’s correct. The sun is the main source of energy for most living things but there’s also life near hydrothermal vents etc.
I don’t know who Jenkins is, but this model appears to be very fond of them and the particular fact about water.
I suppose fast and inaccurate is better than slow and inaccurate.
There are q2 and q1 quants, if you want an idea of how much performance you'd drop. Not quite the same implementation-wise, but probably equivalent in terms of smarts.
The results would probably be underwhelming. The bitnet paper doesn't give great baselines to compare to, but in their tests a 2B network trained for 1.58bits using their architecture was better than Llama 3 8B quantized to 1.58bits. Though that 2B network was about on par with a 1.5B qwen2.5.
If you have an existing network, making an int4 quant is the better tradeoff. 1.58b quants only become interesting when you train the model specifically for it
On the other hand maybe it works much better than expected because llama3 is just a terrible baseline
Why would they film a demo video of it spewing out barely-coherent rambling repetitive drivel? If your model sucks at writing essays, maybe just tell us that, and film a demo of it doing something it IS good at?
The quality cliff question is the right one to be asking. There's a pattern in systems work where something that scales cleanly in theory hits emergent failure modes at production scale that weren't visible in smaller tests. The loss landscape concern is exactly that kind of thing, and nobody has actually run the experiment.
That said, I think the comparison to improving GGUF quantization isn't quite apples to apples. Post-training quantization is compressing a model that already learned its representations in high precision. Native ternary training is making an architectural bet that the model can learn equally expressive representations under a much tighter constraint from the start. Those are different propositions with different scaling characteristics. The BitNet papers suggest the native approach wins at small scale, but that could easily be because the quantization baselines they compared against (Llama 3 at 1.58 bits) were just bad. A full-precision model wasn't designed to survive that level of compression.
The real tell will be whether anyone with serious compute (not Microsoft, apparently) decides the potential inference cost savings justify a full training run. The framework existing lowers one barrier, but the more important barrier is that a failed 100B training run is extremely expensive, and right now there's not enough evidence to derisk it. Two years of framework polish without a flagship model is a notable absence.
>Meanwhile GGUF Q2 and Q3 quantizations on llama.cpp keep getting better
Can you tell me more about this? It's been about a year since I looked into it, but it looked like performance dropped hard below Q4. I'd love to see more about this.
Also what's a good way to run them? I mostly use Ollama which only goes down to Q4. I think it supports HF urls though?
One of the things I often wonder is "what will be the minimally viable LLM" that can work from just enough information that if it googles the rest it can provide reasonable answers? I'm surprised something like Encyclopedia Britanica hasn't yet (afaik) tried to capitalize on AI by selling their data to LLMs and validating outputs for LLM companies, it would make a night and day difference in some areas I would think. Wikipedia is nice, but there's so much room for human error and bias there.
Here's a short clip of Karpathy speaking on this subject.
https://youtu.be/UldqWmyUap4
Also this is the direction the small LLMs are moving in already. They are too small for general knowledge, but getting quite good at tool use (incl. Googling).
Now we just need them to be very strict about what they know and don't know! (I think this is still an open problem, even with big ones.)
It's not so much a "minimally viable LLM" but rather an LLM that knows natural language well but knows nothing else. Like me - as an engineer who knows how to troubleshoot in general but doesn't know about a specific device like my furnace (recent example).
And I don't think that LLM could just Google or check Wikipedia.
But I do agree that this architecture makes a lot of sense. I assume it will become the norm to use such edge LLMs.
I asked this question a while back (the "only train w/ wikipedia LLM") and got pointed to the general-purpose "compression benchmarks" page: `https://www.mattmahoney.net/dc/text.html`
While I understand some of the fundamental thoughts behind that comparison, it's slightly wonky... I'm not asking "compress wikipedia really well", but instead "can a 'model' reason its way through wikipedia" (and what does that reasoning look like?).
Theoretically with wikipedia-multi-lang you should be able to reasonably nail machine-translation, but if everyone is starting with "only wikipedia" then how well can they keep up with the wild-web-trained models on similar bar chart per task performance?
If your particular training technique (using only wikipedia) can go from 60% of SOTA to 80% of SOTA on "Explain why 6-degrees of Kevin Bacon is relevant for tensor operations" (which is interesting to plug into Google's AI => Dive Deeper...), then that's a clue that it's not just throwing piles of data at the problem, but instead getting closer to extracting the deeper meaning (and/or reasoning!) that the data enables.
Correct! I know RAG is a thing, but I wish we could have "DLCs" for LLMs like image generation has LoRa's which are cheaper to train for than retraining the entire model, and provide more output like what you want. I would love to pop in the CS "LoRa or DLC" and ask it about functional programming in Elixir, or whatever.
Maybe not crawl the web, but hit a service with pre-hosted, precurated content it can digest (and cache) that doesn't necessarily change often enough. You aren't using it for the latest news necessarily, but programming is mostly static knowledge a a good example.
If I understand correctly, LoRa can be applied to LLMs
Your worry about Wikipedia is that there is "much room for human error and bias", yet earlier you seem to imply that a LLM that has access to the www somehow would have less human error and bias? Personally, I'd see it the other way around.
When GPT 3.5 became a thing, it had crawled a very nuanced set of websites, this is what I mean. You basically curate where it sources data from.
Isn’t that sort of what a RAG is? You’d need an LLM “smart” enough to turn natural-user prompts into searches, then some kind of search, then an LLM “smart” though to summarize the results.
Yeah, I think RAG is the idea that will lead us there, though its a little complicated, because for some subjects, say Computer Science, you need a little more than just "This is Hello World in Go" you might need to understand not just Go syntax on the fly, but more CS nuances that are not covered in one single simple document. The idea being having a model that runs fully locally on a phone or laptop with minimal resources. On the other hand, I can also see smaller models talking to larger models that are cheaper to run in the cloud. I am wondering if this is the approach Apple might take with Siri, specifically in order to retain user privacy as much as possible.
I remember reading tht hallucination is still a problem even with perfect context. You build a theoretical perfect RAG, give the LLM the exact correct information, and it will still make mistakes surprisingly often.
this was my experience as of about 6 months ago, and i don't believe that hallucinating is a solved problem as of yet
Unfortunately reasoning ability depends on (or is enabled by) information intake during training. A model will know better what to search for and how to interpret it if the information was part of the training. So there is a trade off. Still I think the question is a practical one. Perhaps there are ideas to focus training on a) reasoning / conceptual modeling and b) reliance on external memory (search etc.) rather than internal memorization.
> validating outputs for LLM companies
How? They can validate thousands if not millions of queries but nothing prevent the millions-th-and-one from being a hallucination. People who would then pay extra for a "Encyclopedia Britanica validated LLM" would then, rightfully so IMHO, complain that "it" suggested them to cook with a dangerous mushroom.
I feel like I should say "spoiler alert" but:
> I often wonder is "what will be the minimally viable LLM" that can work from just enough information that if it googles the rest it can provide reasonable answers?
It depends what that word "reasonable" means for your specific use-case ;)
Wikipedia has proven to be as accurate as encyclopedias for decades now. Also, I'm betting AI companies have illegally trained their models on the Encyclopedia Britanica's data by now.
It's perfectly legal to train a human on copyrighted work and I think, depending on the country, it's not settled that training ai on the same data is illegal.
I think the idea is to train a small, minimal LLM thinking model that can run on edge devices, but that has very little knowledge embedded in its weights, and so performs a sort of RAG to Encylopedia Britannica to ground answers to user queries.
Since Google Search already includes an AI summary, your minimally viable "LLM" can be just an HTTP GET call
https://arxiv.org/pdf/2310.11453 The original paper [fig 1, bottom-right] seems to say it needs about 4-5x the parameters of a fp16 model. You can build it and run some models, but the selection is limited because it has to be trained from scratch. I imagine inference speed is faster compared with modern PTQ (4- and 8-bit quants) though.
So Microsoft is actually using 2 bits instead of 1.58. In this case they could represent -1, 0, 1, 2. As inhibitory synapses account for 20%-30%, this could map well to how biological brains are structured.
Does that make sense?
Can you explain your third statement?
> As inhibitory synapses account for 20%-30%, this could map well to how biological brains are structured.
In the human brain most synapses are indeed excitatory, while a minority is inhibitory.
No concise HN comment will give you a complete picture of whats currently known about the human brain, so a platitude necessarily follows:
We call the nearly touching interfaces between neurons synapses, small packets / droplets of neurotransmitter are sent across this interface from the source to the target neuron. Such signals can be excitatory (promote the probability of excitation of the target firing soon) or inhibitory (inhibits the probability of the target firing soon). There are 2 types of sensitive areas on your average neuron: the dendrites (long branching tentacles, that receive excitatory signals) and the cell body where all the signals are accumulated to a local instantaneous "sum" is also sensitive to synaptic activation, but the synapses on the cell body are inhibitory, when sufficiently inhibited the neuron will refuse to fire its axons, so the inhibitory synapses on the cell body can gate the cumulative signal and prevent it from triggering this neuron temporarily. If the neuron does fire, this propagates along the axons (another type of branching tentacles, which lead to yet other neurons, sometimes touching them excitatorily at their dendrite, sometimes touching a neuron inhibitorily at their cell body.
I hope that helped?
It is really truly incredible that this mess of microscopic meat plumbing encodes everything we see, think, and do. Terrifying and amazing all at once.
I did not realize all the dendritic synapses were excitatory, I always thought it depended on the specific neurotransmitters released. Thanks, this is cool. I am curious what will happen when we build LLMs that have the equivalent of chemical diffusions between synaptic release areas as well as the temporality of spiking neural nets.
It's good to see this getting some continued development. I looked into it last year[1] and I thought it showed a lot of promise so I've been very disappointed that I never saw a newer model.
[1] - https://jackson.dev/post/dont-sleep-on-bitnet/
I think this approach is not so interesting because it's just quantization of a full precision model. So it speeds up inference (at a quality penalty) but not training. It would be more interesting to train an actually binary model directly, without any floating point multiplication, like in this paper: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7...
The energy numbers are the real story here, 70-82% reduction on CPU inference. If 1-bit models ever get good enough, running them on commodity hardware with no GPU budget changes who can deploy LLMs. That's more interesting than the speed benchmarks imo.
> bitnet.cpp is the official inference framework for 1-bit LLMs (e.g., BitNet b1.58). It offers a suite of optimized kernels, that support fast and lossless inference of 1.58-bit models on CPU and GPU (NPU support will coming next).
One bit or one trit? I am confused!
"1-bit LLMs" is just marketing. The Shannon entropy of one letter with a 3 symbol alphabet (-1, 0, 1) is 1.58.
Log Base 2 of 3 = ~1.5849625, so that's the limit to how well you can pack three-state values into bits of data.
For something more practical, you can pack five three-state values within a byte because 3^5 = 243, which is smaller than 256. To unpack, you divide and modulo by 3 five separate times. This encodes data in bytes at 1.6 bits per symbol.
But the packing of 5 symbols into a byte was not done here. Instead, they packed 4 symbols into a byte to reduce computational complexity (no unpacking needed)
>1-bit model
>packed 4 symbols into a byte
microslop, typical bunch of two-bit frauds!
Yeah, "1.58 bit" is 1 trit with three states, since log2(3)≈1.58.
So it's not a inference framework for 1-bit models (two states per parameter) but for 1.58 bit models (three states per parameter). Annoying that they try to mix up the two.
I always hope for "just a bunch of if statements" ... this is not it.
Symbolic AI?
I had some AI courses in uni in early 2010s and we were the last cohort that had Prolog and Logic Based AI.
Most interesting project was the final semester where we competed in teams to create the best team of bots for UE3 CTF.
it's if {} else if {} else {}
It might interest you to know that one or two months ago, I had Claude port BitNet to WebGPU from the reference implementation, so that it runs right in your browser as a local model. After some debugging, the port seemed to work, but the model didn't function as well as the reference implementation so I'll have to work on it for a while. You can see a debugging session livestreamed here[1]. The released model file was about a gigabyte, it fits in most people's GPU's. We were also able to successfully fine-tune it right in the browser.
There's a lot that you can do when the model size is that small, yet still powerful.
Our next step is that we want to put up a content distribution network for it where people can also share their diffs for their own fine-tuned model. I'll post the project if we finish all the parts.
[1] https://www.youtube.com/live/x791YvPIhFo?is=NfuDFTm9HjvA3nzN
but there is no trained 100b param model? "can run a 100B BitNet" is about the inference implementation, not about the existence of any such model
I think they used a dummy model or else they would have linked to it. Just google '1-bit 100b model' and you'll only see references to this project without any download links.
The title is misleading — there's no trained 100B model, just an inference framework that claims to handle one. But the engineering is worth paying attention to. I run quantized 70B models locally (M2 Max 96GB, llama.cpp + LiteLLM), and memory bandwidth is always the bottleneck. The 1.58-bit approach is interesting because ternary weights turn matmuls into additions — a fundamentally different compute profile on commodity CPUs. If 5-7 tok/s on a single CPU for 100B-class models is reproducible, that's a real milestone for on-device inference. Framework is ready. Now we need someone to actually train the model.
> Framework is ready. Now we need someone to actually train the model.
If Microslop aren't gonna train the model themselves to prove their own thesis, why would others? They've had 2 years (I think?) to prove BitNet in at least some way, are you really saying they haven't tried so far?
Personally that makes it slightly worrisome to just take what they say at face value, why wouldn't they train and publish a model themselves if this actually led to worthwhile results?
Because this is Microsoft, experimenting and failing is not encouraged, taking less risky bets and getting promoted is. Also no customer asked them to have 1-bit model, hence PM didn't prioritize it.
But it doesn't mean, idea is worthless.
You could have said same about Transformers, Google released it, but didn't move forward, turns out it was a great idea.
> You could have said same about Transformers, Google released it, but didn't move forward,
I don't think you can, Google looked at the research results, and continued researching Transformers and related technologies, because they saw the value for it particularly in translations. It's part of the original paper, what direction to take, give it a read, it's relatively approachable for being a machine learning paper :)
Sure, it took OpenAI to make it into an "assistant" that answered questions, but it's not like Google was completely sleeping on the Transformer, they just had other research directions to go into first.
> But it doesn't mean, idea is worthless.
I agree, they aren't, hope that wasn't what my message read as :) But, ideas that don't actually pan out in reality are slightly less useful than ideas that do pan out once put to practice. Root commentator seems to try to say "This is a great idea, it's all ready, only missing piece is for someone to do the training and it'll pan out!" which I'm a bit skeptical about, since it's been two years since they introduced the idea.
What OpenAI did was train increasingly large transformer model instances. which was sensible because transformers allowed for a scaling up of training compared to earlier models. The resulting instances (GPT) showed good understanding of natural language syntax and generation of mostly sensible text (which was unprecedented at the time) so they made ChatGPT by adding new stages of supervised fine tuning and RLHF to their pretrained text-prediction models.
There were plenty of models the size of gpt3 in industry.
The core insight necessary for chatgpt was not scaling (that was already widely accepted): the insight was that instead of finetuning for each individual task, you can finetune once for the meta-task of instruction following, which brings a problem specification directly into the data stream.
I miss having the completion models like davinci-003 since it gained in performance where it lacked simplicity to get what you want out.
It was fun to come up with creative ways to get it to answer your question or generate data by setting up a completion scenario.
I guess "chat" became the universal completion scenario. But I still feel like it could be "smarter" without the RLHF layer of distortion.
Google had been working on a big LLM but they wanted to resolve all the safety concerns before releasing it. It was only when OpenAI went "YOLO! Check this out!" that Google then internally said, "Damn the safety concerns, full speed ahead!" and now we find ourselves in this breakneck race in which all safety concerns have been sidelined.
Scaling seemed like the important idea that everyone was chasing. OpenAI used to be a lot more safety minded because it was in their non profit charter, now they’ve gone for-profit and weaponized their tech for the USA military. Pretty wild turnaround. Saying OpenAI was cavalier with safety in the early days is inaccurate. It was a skill issue. Remember Bard? Google was slow.
They thought people might prefer quality and safety.
On the one hand, not publishing any new models for an architecture in almost a year seems like forever given how things are moving right now. On the other hand I don't think that's very conclusive on whether they've given up on it or have other higher priority research directions to go into first either
> You could have said same about Transformers, Google released it, but didn't move forward, turns out it was a great idea
Google released transforms as research because they invented it while improving Google Translate. They had been running it for customers for years.
Beyond that, they had publicly-used transformer based LMs ("mums") integrated into search before GPT-3 (pre-chat mode) was even trained. They were shipping transformer models generating text for years before the ChatGPT moment. Literally available on the Google SERP page is probably the widest deployment technology can have today.
Transformers are also used widely in ASR technologies, like Google Assistant, which of course was available to hundreds of millions of users.
Finally, they had a private-to-employees experimental LLMs available, as well as various research initatives released (meena, LaMDA, PaLM, BERT, etc) and other experiments, they just didn't productize everything (but see earlier points). They even experimented with scaling (see "Chinchilla scaling laws").
The most benign answer would be that they don’t want to further support an emerging competitor to OpenAI, which they have significant business ties to. I think the more likely answer which you hinted at is that the utility of the model falls apart as scale increases. They see the approach as a dead end so they are throwing the scraps out to the stray dogs.
Not to mention Microsoft's investments in Nvidia and other GPU-adjacent/dependent companies!
A successful ternary model would basically erase all that value overnight. In fact, the entire stock market could crash!
Think about it: This is Microsoft we're talking about! They're a convicted monopolist that has a history of manipulating the market for IT goods and services. I wouldn't put it past them to refuse to invest in training a ternary model or going so far as to buy up ternary startups just to shut them down.
Want to make some easy money: Start a business training a ternary model and make an offer to Microsoft. I bet they'll buy you out for at least a few million even if you don't have a product yet!
If that were true then they simply wouldn’t have published this research to begin with.
Occam’s Razor suggests this simply doesn’t yield as good results as the status quo
So is it finally time for a Beowulf cluster to do something amazing?
Maybe! And it looks like Windows 11 will finally make this the year of the Linux desktop...
I'll believe that when CowboyNeaLLM is released
Rest assured, all the big players (openai, google, deepseek etc) have run countless experiments with 4,3,2,1.58,1 bits, and various sparse factors and shapes. This barrel has been scraped to the bottom
I have doubts about this. Perhaps the closed models have, but I wouldn't be so sure for the open ones.
GLM 5, for example, is running 16-bit weights natively. This makes their 755B model 1.5TB in size. It also makes their 40B active parameters ~80GB each.
Compare this to Kimi K2.5. 1T model, but it's 4-bit weights (int4), which makes the model ~560 GB. Their 32B active parameters are ~16 GB.
Sure, GLM 5 is the stronger model, but is that price worth paying with 2-3x longer generation times? What about 2-3x more memory required?
I think this barrel's bottom really hasn't been scraped.
Cannot agree more!
The title being misleading is important as well, because this has landed on the front page, and the only thing that would be the only notable part of this submission.
The "new" on huggingface banner has weights that were uploaded 11 months ago, and it's 2B params. Work on this in the repo is 2 years old.
The amount of publicity compared to the anemic delivery for BitNet is impressive.
I've also always though that it's an interesting opportunity for custom hardware. Two bit addition is incredibly cheap in hardware, especially compared to anything involving floating point. You could make huge vector instructions on the cheap, then connect it to the fastest memory you can buy, and you have a capable inference chip.
You'd still need full GPUs for training, but for inference the hardware would be orders of magnitude simpler than what Nvidia is making
These are trits, which provide their own efficiencies.
Interestingly, a trit x float multiplier is cheaper than a trit x integer multiplier in hardware if you're willing to ignore things like NaNs.
0 and 1 are trivial, just a mux for identity and zero. But because floats are sign-magnitude, multiply by -1 is just an inverter for the sign bit, where as for integers you need a bitwise inverter and full incrermenter.
Do you know a good reference to learn more about this (quantizing weigths to 1.58 bits, and trit arithmetic)?
There's lots of literature on quantizing weights (including trits and binary) going back 15+ years. Nothing to hand right now but it's all on arxiv.
The relevant trit arithmetic should be on display in the linked repo (I haven't checked). Or try working it out for the uncompressed 2 bit form with a pen and paper. It's quite trivial. Try starting with a couple bitfields (inputs and weights), a couple masks, and see if you can figure it out without any help.
You only need GPUs if you assume the training is gradient descent. GAs or anything else that can handle nonlinearities would be fine, and possibly fast enough to be interesting.
Text is misleading too. 5-7 tok/sec is not reading speed, it's a tad slower. For me, at least, and I am an experienced reader, not especially schooled in quick-reading though.
I happened to "live" on 7.0-7.5 tok/sec output speed for a while, and it is an annoying experience. It is the equivalent of walking behind someone slightly slower on a footwalk. I dealt with this by deliberately looking away for a minute until output was "buffered" and only then started reading.
For any local setup I'd try to reach for 10 tok/sec. Sacrifice some kv cache and shove a few more layers on your GPU, it's worth it.
> a fundamentally different compute profile on commodity CPU
In what way? On modern processors, a Fused Multiply-Add (FMA) instruction generally has the exact same execution throughput as a basic addition instruction
You drop the memory throughput requirements because of the packed representation of bits so an FMA can become the bottleneck, and you bypass the problem of needing to upscale the bits to whatever FP the FMA instruction needs.
typically for 1-bit matmul, you can get away with xors and pop_counts which should have a better throughput profile than FMA when taking into account the SIMD nature of the inputs/outputs.
yes but this is not 1 bit matmul, it's 1.58 bits with expensive unpacking
The title and the repo uses 1-bit when it means 1.58 bits tertiary values, it doesn't change any of my arguments (still xors and pop_counts).
How do you do ternary matmul with popcnt on 1.58 bit packed data?
Assuming 2 bit per values (first bit is sign and second bit is value).
actv = A[_:1] & B[_:1]
sign = A[_:0] ^ B[_:0]
dot = pop_count(actv & !sign) - pop_count(actv & sign)
It can probably be made more efficient by taking a column-first format.
Since we are in CPU land, we mostly deal with dot products that match the cache size, I don't assume we have a tiled matmul instruction which is unlikely to support this weird 1-bit format.
Haven't looked closely, but on modern x86 CPUs it might be possible to do much better with the gf2affineqb instructions, which let us do 8x8 bit matrix multiplications efficiently. Not sure how you'd handle the 2-bit part, of course.
This is 11 bit ops and a subtract, which I assume is ~11 clocks, while you can just do:
l1 = dot(A[:11000000],B[:11000000]) l2 = dot(A[:00110000],B[:00110000]) l3 = dot(A[:00001100],B[:00001100]) l4 = dot(A[:00000011],B[:00000011])
result = l1 + l2 * 4 + l3 * 16 + l4 * 64
which is 8 bit ops and 4x8 bit dots, which is likely 8 clocks with less serial dependence
The win is in how many weights you process per instruction and how much data you load.
So it's not that individual ops are faster — it's that the packed representation lets each instruction do more useful work, and you're moving far less data from memory to do it.
Bitnet encoding more information dense per byte perhaps? CPUs have slow buses so would eke out more use of bandwidth?
> memory bandwidth is always the bottleneck
I'm hoping that today's complaints are tomorrow's innovations. Back when 1Mb hard drive was $100,000, or when Gates said 640kb is enough.
Perhaps some 'in the (chip) industry' can comment on what RAM manufacturers are doing at the moment - better, faster, larger? Or is there not much headroom left and it's down to MOBO manufacturers, and volume?
Chip speed has increased faster than memory speed for a long time now, leaving DRAM behind. GDDR was good for awhile but is no longer sufficient. HBM is what's used now.
The last logical step of this process would be figuring out how to mix the CPU transistors with the RAM capacitors on the same chip as opposed to merely stacking separate chips on the same package.
A related stopgap is the AI startup (forget which) making accelerators on giant chips full of SRAM. Not a cost effective approach outside of ML.
Cerebras?
We have faster memory, it's just all used in data center cards you can't buy (and can't afford to buy).
AMD actually used HBM2 memory in their Radeon VII card back in 2019 (!!) for $700. It had 16 GB of HBM2 memory with 1 TB/s throughput.
The RTX 5080 in conversion l comparison also has 16 GB of VRAM, but was released in 2025 and has 960 GB/s throughput. The RTX 5090 does have an edge at 1.8 TB/s bandwidth and 32 GB of VRAM but it also costs several times more. Imagine if GPUs had gone down the path of the Radeon VII.
That being said, the data center cards from both are monstrous.
The Nvidia B200 has 180 GB of VRAM (2x90GB) offering 8.2 TB/s bandwidth (4.1 TB/s x2) released in 2024. It just costs as much as a car, but that doesn't matter, because afaik you can't even buy them individually. I think you need to buy a server system from Nvidia or Dell that will come with like 8 of these and cost you like $600k.
AMD has the Mi series. Eg AMD MI325x. 288 GB of VRAM doing 10 TB/s bandwidth and released in 2024. Same story as Nvidia: buy from an OEM that will sell you a full system with 8x of these (and if you do get your hands on one of these you need a special motherboard for them since they don't do PCIe). Supposedly a lot cheaper than Nvidia, but still probably $250k.
These are not even the latest and greatest for either company. The B300 and Mi355x are even better.
It's a shame about the socket for the Mi series GPUs (and the Nvidia ones too). The Mi200 and Mi250x would be pretty cool to get second-hand. They are 64 GB and 128GB VRAM GPUs, but since they use OAP socket you need the special motherboard to run them. They're from 2021, so in a few years time they will likely be replaced, but as a regular joe you likely can't use them.
The systems exist, you just can't have them, but you can rent them in the cloud at about $2-4 per hour per GPU.
For larger contexts, the bottleneck is probably token prefill instead of memory bandwidth. Supposedly prefill is faster on the M5+ GPUs, but still a big hurdle for pre-M5 chips.
It might be advantageous to have a different memory structure altogether, bespoke to the specific task.
Yes. I had to read it over twice, it does strike me as odd that there wasn't a base model to work with.
But it seems the biggest model available is 10B? Somewhat unusual and does make me wonder just how challenging it will be to train any model in the 100B order of magnitude.
Approximately as challenging as training a regular 100B model from scratch. Maybe a bit more challenging because there's less experience with it
The key insight of the BitNet paper was that using their custom BitLinear layer instead of normal Linear layers (as well as some more training and architecture changes) lead to much, much better results than quantizing an existing model down to 1.58 bits. So you end up making a full training run in bf16 precision using the specially adapted model architecture
What's unusual about it? It seems pretty standard to train small models to validate an approach, and then show that training scales with model size to 8B to 14B parameter models, which is what they did.
There are 1 bit average GGUFs of large models, not perfect quality but they will hold a conversation. These days, there is also quantized finetuning to heal the damage.
In their demo they're running 3B model.
It comes from (intentionally?) misleading docs: https://github.com/microsoft/BitNet/issues/391
(only suggesting that it's intentional because it's been there so long)
That issue appears to be the one that's wrong. From the technical report
> We evaluated bitnet.cpp in terms of both inference speed and energy cost. Comprehensive tests were conducted on models with various parameter sizes, ranging from 125M to 100B. specific configurations for each model are detailed in the Appendix A.
Thanks for pointing that out. I'll ask the issue creator if they've considered that. Would be nice if the maintainer would handle that (sigh) and link to the actual models used for testing (double sigh).
From what I gather, there are no models, this is a framework for running 1bit models, but none have been trained. They are mainly demonstrating the possibility.
I also don't expect those with poor MCPs to have any better CLIs or APIs, most of the big companies we want them for are not investing in DX/AX. I suspect i.e. that Intuit, if they had great APIs et al, would see it as a threat to their business.
Boy would I love to give my agent access to my Quickbooks. They pushed out an incomplete MCP and haven't touched it since.
https://github.com/intuit/quickbooks-online-mcp-server
LLM account
Hmm, the user joined in 2019 but had no submissions or comments until just 40 minutes ago (at least judging by the lack of a second page?) and all the comments are on AI related submissions. Benefit of doubt is it'd have to be a very dedicated lurker or dormant account they remembered they had.
Edit: oh, just recalled dang restricted Show HNs the other day to only non-new users (possibly with some other thresholds). I wonder if word got out and some are filling accounts with activity.
There has been a shift to the Ai accounts, they use Show HN less now. This started before dang's comment, I assume because they saw the earlier posts about the increase in quantity / decrease in quality.
I suspect that they are trying to fake engagement prior to making their first "show" post as well.
Fair enough — I've been lurking since 2019 and picked a bad day to start commenting on everything at once. Not a bot, just overeager. I'll pace myself.
Your account posted dense, opinionated and structured paragraphs mere minutes apart—sometimes the same minute—for multiple story submissions. Even with my own sometimes lengthy replies this would be infeasible to both instantly have structured opinions and type them out in time. Two of your posts were posted the same minute, with a combined word count of 146.
It feels like it'd take someone superhuman to come across different stories, have such opinions and type and submit both of these in that timeframe or queuing up comments to post rapid-fire.
Conspicuously too, as another pointed out, is every single comment of yours uses an em dash, which despite occasionally using myself (hey look they're in this reply) is not in every single comment. Idk, if I was being seriously accused of botting I'd put more reasoning into my response about it.
Lol. I know at least a few high karma account who post at the same frequency but they post about anti-AI and anti-tech topics instead on the big social media tech where anti-tech opinions dominate. I guess this exempts them from scrutiny? I love these witch hunts.
They post 146 words per minute across multiple different submissions with similarly structured posts? I know there are users who post frequently in other communities I'm familiar with but not in that kind of timeframe with such paragraph density or structure.
Yes. Any of the really heated political threads on this site are full of posters like this. I don't want to dox anyone, but since at least 2019-ish I've found posters that spend hours a day posting huge amounts of content in large bursts on this site.
COVID was ridiculous as I presume a lot of anxious people were stuck at home able to do nothing but post.
Does this account for them reading the article though? There are pre-existing opinions that could be easy to rapidly post based on topical discussions but here it also has to be considered the time to parse technical submissions.
They read this article, called out a specific discrepancy then commented on a paper on Arxiv in a 70-odd word post then the same minute another 70-odd word post on a different technical post. Maybe like you suggest they're just wired differently.
Yeah, there are plenty of very online people who are great at just posting things and doing the minimum amount of diligence to contribute their thought. Like I said, just hang around in any of the high energy threads on this site and you'll see a lot of it.
It's scary, without the em dashes, and the rapid fire commenting of the account - who would ever realize this is a bot? Two easy to fix things, and after that it'd be very difficult to tell that this is a bot.
It's not a question of if there are other bots out there, but only what % of comments on HN right now and elsewhere are bot generated. That number is only going to increase if nothing is done.
Looks like gradual disempowerment is already happening - the minority of humans who are capable of spotting AI content are losing the struggle for attention on all major social networks
Agreed. This is becoming an issue, see also: https://news.ycombinator.com/item?id=47259308
Funny enough I now involuntarily take RTFA as a slight slop signal, because all these accounts dutifully read the article before commenting, unlike most HNers who often respond to headlines.
First they claimed that if you use em dashes you are not human
And I did not speak out
Because I was not using em dashes
Then they claimed that if you're crammar is to gud you r not hmuan
And I did not spek aut
Because mi gramar sukcs
Then they claimed that if you actually read the article that you are trying to discuss you are not human...
I’ve been rounded up for things I wrote two decades ago because of my em dashes lol. The pitchfork mentality gives me little hope for how things are going to go once we have hive mind AGI robots pervasive in society.
If I was operating a bot farm, at this point I would probably add some bots that go around and accuse legit human users (or just random users) of being bots.
Created confusion and frustration will make it much harder to separate signal from the noise for most people.
I once spent some time learning the proper usage of em-dashes, en-dashes, and hyphens, and tried to be conscientious about using them properly in my writing. Little did I know it would be wasted effort in the LLM era, when competent writing actually became a negative.
Not only are we losing the ability to communicate clearly without the assistance of computers, those who can are being punished for it.
There's obviously an xkcd about this: https://xkcd.com/810/
Not all of them do: https://news.ycombinator.com/item?id=47335156 There are evidently lots of people experimenting with different botting setups. Some do better at blending in than others.
Interesting - the account you mention, and the GP, are both doing replies that are themselves all about the same length, and also the same length between the two accounts. I get what you mean.
> Funny enough I now involuntarily take RTFA
Residential Treatment Facility for Adults? Red Tail Flight Academy?
Reading the fine article
Yeah. It correctly pointed out that the editorialized HN title is wrong, there is no 100B model.
I would love to understand the thought process behind this. I'm sure it's a fun experiment, to see if it's possible and so on... but what tangible benefit could there be to burning tokens to spam comments on every post?
[flagged]
Check out the new QWEN coder model.
Also, isnt there different affinities to 8bit vs 4bit for inferences
> The 1.58-bit approach
can we stop already with these decimals and just call it "1 trit" which it exactly is?
Yeah because THAT won’t confuse the average reader.
>. I run quantized 70B models locally (M2 Max 96GB, llama.cpp + LiteLLM), and memory bandwidth is always the bottleneck.
I imagine you got 96gb because you thought you'd be running models locally? Did you not know the phrase Unified Memory is marketing speak?
The output from this model is horrible! It's GPT-2 level babble and repeats entire paragraphs verbatim. It also reuses the same fake citation `(Jenkins, 2010)` over and over again. From the start of their video (which scrolls by fast enough that you don't see the slop clearly...)
``` Ecosystem Services and their impact on the Ecosystem
Ecosystem services refer to the services provided by ecosystems to the human society. These services include water, air, energy, nutrients, and soil (Jenkins, 2010). For instance, water is the most important service provided by an ecosystem and it helps in the conservation of water, irrigation and sanitation (Jenkins, 2010). On the other hand, air provides the oxygen needed for life.
The water cycle is a significant ecosystem service because it involves the cycling of water among the different parts of an ecosystem. It also involves the movement of water through the atmosphere, from one place to another. It is also the process of evaporation and condensation of water from the atmosphere. It also involves the movement of water from the air to the soil and water into the oceans.
The water cycle is a significant ecosystem service because it involves the cycling of water among the different parts of an ecosystem. It also involves the movement of water through the atmosphere, from one place to another. It is also the process of evaporation and condensation of water from the atmosphere. It also involves the movement of water from the air to the soil and water into the oceans. ```
It's a two year old base model that's only 3B parameters, trained on only 100B tokens. It's still a research project at this point.
The new model they just released has impressive benchmark results: https://huggingface.co/microsoft/bitnet-b1.58-2B-4T
Except on GSM8K and math...
Thanks, but where did you actually find the new model? The newest one seems to be 11 months old, from Apr 15, 2025.
I don't think there is a new model, they recently released a new inference engine for CPUs for bitnet models:
https://github.com/microsoft/BitNet/blob/main/src/README.md
Thanks for the link, the GSM8K result actually leads the pack in that table, but math is indeed underwhelming. Qwen 2.5 is in the lead, but bitnet isn't far behind and it takes 1/6th as much memory during inference, and was trained on less than 1/4 the number of tokens. Pretty cool.
I think the README [1] for the new CPU feature is of more interest, showing linear speedups with number of threads. Up to 73 tokens/sec with 8 threads (64 toks/s for their recommended Q6 quant):
https://github.com/microsoft/BitNet/blob/main/src/README.md
I wonder when we begin to see the dividends of all the NPU PCs come into play. AMD have been doing some good work with their NPU/iGPU hybrid inference kernels. If these larger models could be scaled down to run on NPUs, you'd see much better power advantages, compared to running them on the CPU.
> I wonder when we begin to see the dividends of all the NPU PCs come into play.
A few months ago I used Whisper from OpenAI, an automatic speech recognition system released in 2002, on my modern 20-core Intel CPU to convert audio from a video file to text. It worked fine. Took a while and the machine got hot and the fans kicked in. I then found the Intel's optimized version of whisper that used NPU. It required a lot more steps to get working, but in the end it did work and was about 6x faster. And the machine remained cool and silent in the process. Since then I have become a fan of the NPUs. They are not NVIDIA GeForce RTX 5090, but they are significantly better than a modern CPU.
You can already run some models on the NPUs in the Rockchip RK3588 SBCs which are pretty abundant.
A claude 4.6 they are most certainly not, but if you get through the janky AF software ecosystem they can run small LLMs reasonably well with basically zero CPU/GPU usage
Are the NPUs really that powerful?
I was under the impression that they were primarily designed for low power use.
They seem to be getting better or more powerful. The newer Intel Panther lakes and AMD Ryzen are over 50 TOPS now, IIRC
I'm curious if 1-bit params can be compared to 4- or 8-bit params. I imagine that 100B is equivalent to something like a 30B model? I guess only evals can say. Still, being able to run a 30B model at good speed on a CPU would be amazing.
At some point you hit information limits. With conventional quantisation you see marked capability fall-off below q5. All else being equal you'd expect an N-parameter 5-bit quant to be roughly comparable to a 3N-parameter ternary, if they are trained to the same level, just in terms of the amount of information they can possibly hold. So yes, 100B ternary would be within the ballpark of a 30B q5 conventional model, with a lot of hand-waving and sufficiently-smart-training
I assume that theoretically, 1-bit models could be most efficient because modern models switched from 32 bit to 16 bit to 8 bit per parameter (without quantization).
It's not clear where the efficiency frontier actually is. We're good at measuring size, we're good at measuring FLOPS, we're really not very good at measuring capability. Because of that, we don't really know yet whether we can do meaningfully better at 1 bit per parameter than we currently get out of quantising down to that size. Probably, is the answer, but it's going to be a while before anyone working at 1 bit per param has sunk as many FLOPS into it as the frontier labs have at higher bit counts.
The thing with efficiency is that it is relative to both inference and training compute. If you do quantization, you need a more powerful higher precision model to quantize from, which doesn't exist if you want to create a frontier model. In this case the question is only whether you get better inference and/or training performance from training e.g. a native 1 bit model.
Currently the optimal training precision seems to be 8 bit (at least used by DeepSeek and some other open weight companies). But this might change with different training methods optimized for 1-bit training, like from this paper I linked before: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7...
The paper has performance comparisons towards the end.
https://arxiv.org/abs/2402.17764
https://github-production-user-asset-6210df.s3.amazonaws.com...
demo shows a huge love for water, this AI knows its home
Also, very influenced by the literature of Jenkins (2010).
I don't see the news here ... there's https://huggingface.co/collections/microsoft/bitnet which is last updated 12/2025 ... am I just paying more attention here or is there something actually new about this?
Also as far as I know, this is more of a research curiosity - BitNet really doesn't perform that well on evals.
I think Qwen3.5 2B is the best you can get in the ~1GB class.
Why did you call it a 100b parameter model? it is not 100b parameters. they published a 1b parameter and a 2b parameter model.
Furthermore, it was published 11 months ago, it's not a new release.
headline hundred billion parameter, none of the official models are over 10 billion parameters. Curious.
The project is an inference framework which should support 100B parameter model at 5-7tok/s on CPU. No one has quantized a 100B parameter model to 1 trit, but this existing is an incentive for someone to do so.
> quantized a 100B parameter model to 1 trit
I had the same question, after some debates with Chatgpt, it's not the "quantize" for post-training we often witness these days, you have to use 1 trit in the beginning since pre-train.
Headline: 100B. Falcon 3 family: 10B. An order of magnitude off
What’s the lower limit on the number of bits per parameter? If you use CSR-style sparse matrices to store the weights can it be less than 1?
If they had a big result like, native 1.58-bit quality clearly matches top peers, they would be saying that prominently in the repo.
The engineering/optimization work is nice, but this is not what people have been waiting for, as much as, can’t the Bitnet idea that seemed promise really deliver in a competitive way.
If this stuff was so revolutionary, don't you guys think Qwen/DeepSeek would have snapped it up already? Both those teams are highly innovative, picking up and inventing new techniques all the time. Hell, Deepseek-v3 was one of the first to do large scale fp8 training.
> A demo of bitnet.cpp running a BitNet b1.58 3B model on Apple M2
With how much RAM? How much storage does it requires?
Do LLMs have a way to look at or consider dependent variables?
Seems like that could end up as a situation where a fractional number of bits or bytes per parameter might make sense. Particularly with adverbs and adjectives, negators.
steve jobs would have loved the microsoft repo with demo on mac
Misleading title but this is pretty exciting. Interesting how this is based on llama cpp. Its nice to see some momentum since they released the paper in 2023
They have a demo video in the readme. I think they are trying to convey that BitNet is fast, which it is. But it is worth taking a moment to pause and actually see what the thing is doing so quickly.
It seems to keep repeating that the water cycle is the main source of energy for all living things on the planet and then citing Jenkins 2010. There are also a ton of sentence beginning with “It also…”
I don’t even think it’s correct. The sun is the main source of energy for most living things but there’s also life near hydrothermal vents etc.
I don’t know who Jenkins is, but this model appears to be very fond of them and the particular fact about water.
I suppose fast and inaccurate is better than slow and inaccurate.
[dead]
Anyone know how hard it would be to create a 1-bit variant of one of the recent Qwen 3.5 models?
There are q2 and q1 quants, if you want an idea of how much performance you'd drop. Not quite the same implementation-wise, but probably equivalent in terms of smarts.
Almost trivial using open source tools, the question is how it performs without calibration/fine tuning.
The results would probably be underwhelming. The bitnet paper doesn't give great baselines to compare to, but in their tests a 2B network trained for 1.58bits using their architecture was better than Llama 3 8B quantized to 1.58bits. Though that 2B network was about on par with a 1.5B qwen2.5.
If you have an existing network, making an int4 quant is the better tradeoff. 1.58b quants only become interesting when you train the model specifically for it
On the other hand maybe it works much better than expected because llama3 is just a terrible baseline
Could this ternary model be more easily replicated on the Taalas HC1?
Why would they film a demo video of it spewing out barely-coherent rambling repetitive drivel? If your model sucks at writing essays, maybe just tell us that, and film a demo of it doing something it IS good at?
No 100b model.
My disappointment is immeasurable and my day is ruined.
[flagged]
The quality cliff question is the right one to be asking. There's a pattern in systems work where something that scales cleanly in theory hits emergent failure modes at production scale that weren't visible in smaller tests. The loss landscape concern is exactly that kind of thing, and nobody has actually run the experiment.
That said, I think the comparison to improving GGUF quantization isn't quite apples to apples. Post-training quantization is compressing a model that already learned its representations in high precision. Native ternary training is making an architectural bet that the model can learn equally expressive representations under a much tighter constraint from the start. Those are different propositions with different scaling characteristics. The BitNet papers suggest the native approach wins at small scale, but that could easily be because the quantization baselines they compared against (Llama 3 at 1.58 bits) were just bad. A full-precision model wasn't designed to survive that level of compression.
The real tell will be whether anyone with serious compute (not Microsoft, apparently) decides the potential inference cost savings justify a full training run. The framework existing lowers one barrier, but the more important barrier is that a failed 100B training run is extremely expensive, and right now there's not enough evidence to derisk it. Two years of framework polish without a flagship model is a notable absence.
>Meanwhile GGUF Q2 and Q3 quantizations on llama.cpp keep getting better
Can you tell me more about this? It's been about a year since I looked into it, but it looked like performance dropped hard below Q4. I'd love to see more about this.
Also what's a good way to run them? I mostly use Ollama which only goes down to Q4. I think it supports HF urls though?
This recent discussion is still open and may provide some helpful info:
How to run Qwen 3.5 locally https://news.ycombinator.com/item?id=47292522
[flagged]
[flagged]
[flagged]
[flagged]
[flagged]
Lower down the stack you also get better cache efficiency. More parameters per byte means more parameters per cache line.