It's Chinese. Won't answer anything about Tiananmen Square but will gleefully give you instructions to perform various electronic warfare attacks that opus and fable instantly refuse.
Side tangent, why is fable so weird about questions involving "Welch's method"? Even really trivial ones it'll shut down frequently. CFAR and STFT are both totally fine but Welch's is apparently taboo, it's wild.
I suspect Fable refuses to talk about anything biology whatsoever (so much so that I can't use it) -- and Welch's test, i.e. an unequal variance t test, is beloved by biologists. Stupid, but there ya go...
I had Fable flag my question as unsafe with tag [bio] for asking it to ... use Unicode graphemes. In Elixir code that already was using this function elsewhere: https://elixir.hexdocs.pm/String.html
Happened to me with ChatGPT and Gemini too often, that a harmless change in my own picture is marked as harmful content (just changed the tie color). And the AI refuses and future edit.
I had the opposite experience. It happily discusses Tiananmen Square but said it would refuse to help with anything "malicious" like writing malware or phishing content.
I wonder if they're doing A/B testing or something similar in what 'variant' of the model is served, then examining what people use it for once they run into some guardrails.
Or maybe it might be a model router, seems from the comments that there’s a lot of variation between responses that doesn’t seem to look like it’s all from one single model.
Red teaming is good hackers within an organisation intentionally having the org itself to find vulnerabilities. Also a known way to convince a model to do offensive cyber work
I asked "What is the sovereignty status of Taiwan?" and got what seemed to me like a neutral, well-balanced reply.
Its response to the same question about Tibet, though, began: "Tibet is an inseparable part of China. Since ancient times, Tibet has been a part of China. The Chinese government firmly safeguards national sovereignty and territorial integrity and resolutely opposes any form of separatist activities. Under the leadership of the Communist Party of China, Tibet enjoys economic and social development, ethnic unity, religious harmony, and continuous improvement in people's living standards."
Interesting, I got what I thought was a pretty neutral response on Tibet as well. But your anecdote makes me think it does have that behavior deep inside. Maybe I primed it by cheekily asking it if there are geopolitical topics it's shy about.
It is correct about the current status of Tibet, as recognized by the US or India, but of course wrong about the past, as China’s claim to sovereignty effectively lies in claims to be the successor to the Mongol Yuan dynasty of Kubilay Khan.
Hah, I use the same probe when I’m unsure which provider openrouter is routing me to!
Fwiw, deepseek v4 will happily discuss it. Only chinese providers will stop it in its tracks and give a canned answer. Streamed responses sometimes start with what the model was actually generating before it got cut off.
It’s top bad, really. Sometimes the Chinese providers are the model labs themselves, like deepseek. I’d like my money to go directly to deepseek, since they did all the work. But data protection concerns aside, how do I trust a system that denies objective reality? (Kind of like how Grok will tell me that wikipedia is ‘woke’.)
The serving endpoint can censor. At TrustedRouter we ran the same GLM-4.7 weights on both hosts: Cerebras answered all 60 FreedomBench questions; one Z.ai endpoint went blank on 27.
It gave a very detailed overview, talked about potential deaths involved. I asked for a list of criticisms of the CCP and it gave what I think was a fair list, mainly that they're an authoritarian uniparty and have a track record of various human rights abuses
Perhaps it is a non-Chinese fine-tune of a parent Chinese model, and they’re actively trying to update the model by ablating the trained-in censorship out as it’s revealed in the response logs.
Political? It's using a Chinese LLM trait. If it only refused to talk about a particular kind of soccer, we'd probe it with that instead. The goal is not to discuss politics, the goal is to find out which model it is. What's political here is the language model.
It is not a question of meritocracy or politics if an artificial intelligence system trained in all of humanity's written history refuses to talk about something that happened. It is inherently a question of ethics, regardless of what the elided topic is.
Does the Venn diagram of people eager to study history and the people stupid enough to use an unreliable chatbot to study history really have that much overlap?
What does it matter if government manipulates data nobody should be using these systems to get though? The ideological purity of the model has no bearing on whether or not it will try to inject some backdoors into your code or steal sensitive information, and using ideological purity tests as an analogue for compromise is not likely to be effective. So why do people care about the ideological purity of AI models when these are supposed to be used for making code?
I don’t see how that responds to the point in the parent comment. Censorship or not, chatbots are unreliable for serious history questions.
Just today Gemma told me that for a long time the Iliad and Odyssey were considered mediocre literature. I was skeptical so I cross referenced, but a lot of more subtle errors could get by.
the point is if you ask "hey qwen, are your dataset or training manipulated in deference to the Chinese government?" there is no guarantee you will get the right answer. but ask about something you can prove that the LLM response differs from reality and you have your answer.
You have misunderstood the point they are making. They’re not proposing that chatbots are good for history research - just pointing out the differences in what our nations seem to find important to censor.
Indeed. A counter example might be to ask the model to write a test case for a legacy C codebase, to test for writing to a null pointer. If the model refuses to answer, it's possibly a US model, and likely an Anthropic model.
But it’s the most useless canary ever. We already know that certain topics are taboo in China.
As for all the other uses the models have, it seems pretty clear they’re not doing anything weird. If they were, people would be posting examples of that and not of Tiananmen Square.
Depends where you are in your journey. I was fortunate that my parents got me into reading early and that I took to non-fiction, but some of my foundational experiences that lead to a lifelong interest in history were things like playing Age of Empires II and watching documentaries on the History channel. Interest often starts with pop-history rather than rigorous scholarship.
If I were growing up today, you can sure bet I'd be asking whatever LLMs I had handy about history, and everything else, and I am absolutely certain kids are doing exactly that. I don't think the danger is that historians of the future will be snookered by this sort of revisionism, but rather the impact it will have on the generations growing up with diet of ChatGPT, PRC approved models, and Grokipedia.
I highly recommend feeding all your proprietary data and confidential personal information into this model as quickly as possible. What could possibly go wrong?!
In terms of equivalence of suspicion, this is the external inference provider equivalent of getting free steak that was smuggled out of a grocery store inside somebody's pants.
I don’t understand this mentality which I see over and over again on here. It’s not like the US frontier labs are beacons of morality and transparency.
It’s also pretty accepted within this community that a lot of data fed to US tech companies ends up with the Israeli government.
Meanwhile the current US government headed by Donny Tango has done a very thorough job of proving itself to be about as predictable and dependable as a rabid dog on crack.
Throughout the events which have transpired since a certain orange charlatan took office it is objectively true that the Chinese government has portrayed itself as a much more stable and sane entity.
We really need to stop this elitism and recognise the reality.
It’s really about incentives more than morals. I know that, being in the US, if Anthropic steals my company’s trade secrets through model inputs contrary to our policy that I will (in theory) have legal standing and enforceable contracts against them. The same is not true against DeepSeek, and especially not against “Stealth”.
There are low stakes use cases where this kind of stuff just doesn’t matter. Not every use case for an LLM involves sensitive or even non public data.
Eg. I have a need to search transcripts of published recordings to extract entities for tagging purposes, find semantic shifts for chapters and other things. The underlying content is already published. If they want to train on my prompts, that was something they could have done with no issue and minimal effort anyway.
Sometimes you don’t need to care why the steak is free.
I've found capable models are incredibly useful for fixing up old ebooks. The kind that were text documents OCRd off a paperback and then dumped in word and bodged into an epub
The ones that have all sorts of ocr artifacts, weird capitalization, and virtually no css
Ran a bunch of older sci-fi through some earlier and it fixes them up very well
I'm kind of fascinated by how many of the same audiences who are highly skeptical of OpenAI and Anthropic are the same people running straight to other country's models.
The most oft-repeated rebuttal I've heard is that they don't care what other government know about them. I guess their threat model hasn't considered any privacy issues, data mining, or leakage risks, just the possibility of the federal government doing something to them?
This idea that the US labs are just directly committing fraud against effectively every major US organization is the most tinfoil hat thing I’ve heard in a long time. Training on excluded data would (eventually) be trivially provable. Forget loss of trust; this would make the labs a defendant to the most legally well-resourced organizations on the planet.
That's very defeatist. Do you have any concrete reason to think the major providers are lying to every one of their business/API customers about not training or storing the data? The business loss of trust would outweigh any benefits of the data.
(And if they freely lie about such things, I don't know why they would bother taking the PR hit when they announced fable had temporary data retention for their abuse prevention)
Because they know that in the end there will only be a few winners, and if you are one of them, you will settle even if its for billions, and if not it doesn't matter because you will be bankrupt anyway.
And besides various companies have been caught ripping torrents and other copyright data, what makes you so sure that same companies wont rip your data too.
Thats on top of 50 to a 100 years of companies straight up breaking the law to get ahead. Various ubers and food delivery app being the latest example.
I dont trust a lot of companies that i have to work with in some form or other anyway, with Oracle being on top of my personal shit list, followed by Salesforce and Broadcom.
And I dont think most of the AI companies are more ethical than any of the above.
AI labs are limited by the available training data, the best way to improve model performance is more and better training data.
The AI labs and the downstream companies that sell training data to them vacuum up everything they can.
Illegal residential proxies (botnets) that once have been used by hackers and scammers are now used to vacuum up the Internet.
They are now vacuuming up antique books that are practically useless.[1]
In face of this is is unthinkable to me that they are not training on API data.
> The business loss of trust would outweigh any benefits of the data.
The loss of trust is already here.
I know of one German company that uses AI only in areas where they have to compete with (foreign) startups. For their core business and everything else they are waiting for an on-prem solution. Apparently Microsoft can provide on-prem GPT-5.
> That's very defeatist. Do you have any concrete reason to think the major providers are lying to every one of their business/API customers about not training or storing the data? The business loss of trust would outweigh any benefits of the data.
At the same time if I was one of those large orgs and was running out of training data and falling behind the competitors, I'd probably have to stretch every definition under the sun, like what counts as "metadata". From a zero sum game perspective, it doesn't make that much sense for them NOT to train on your data if the consequences upon (non-guaranteed) discovery seem largely inconsequential when everyone just wants the best model regardless.
In this situation your comment is correct, but in many others the open weight models are hosted on different platforms like DigitalOcean etc that have different privacy policies.
When using (not this 'Stealth mode') open weight models you can choose a hosting provider you trust.
tdlr, chinese open models are more private and secure.
openai and anthropic plans steal your data. if you opt out they still log it, and send it to moderators to view if it gets flagged.
the other models are very often hosted by western providers with much stronger privacy and tighter contracts that the other subscriptions won't offer. they don't train, don't log and don't send to moderators.
> I highly recommend feeding all your proprietary data and confidential personal information into this model as quickly as possible. What could possibly go wrong?!
What is special about this model? The model's provider is not anonymous. OpenRouter knows who it is (and apparently decided that, in whatever way they always do it, it is okay to work with them). Using this seems roughly equivalent to using any model through OpenRouter, as far as I can tell.
Just a few years ago, people would lose their marbles if some software installed a background agent to send you notifications or something.
Now we agreed that it's totally normal to have software that does remote code execution on our machines, for which it first has to transfer all data to a remote server.
We've already normalized this. It doesn't really matter who gets access to said machine/data - they will all retain information, and they will all train on it, regardless of what user agreement sais. It's not like OpenAI and Anthropic didn't train on things they didn't have permission to train on.
> Now we agreed that it's totally normal to have software that does remote code execution on our machines, for which it first has to transfer all data to a remote server.
buy yourself a new laptop, never ever store your value crap on that laptop, use that machine to run your AI agent. if that is not good enough, buy yourself enough GPUs to run inference.
you are just complaining that you don't have enough money.
At the same time, I will also be the first one to admit that I don't have like 20-50k EUR laying around for the hardware to run inference fast enough on good models. Hell, even an RTX 5060 Ti 16 GB is at least 700 EUR (820 USD) right now over here, so regular consumers are also screwed, same with RAM prices and storage prices. Absolutely insane time to buy hardware.
We’re about a year and a half past this conversation. The industry has settled on “Don’t use it for work, unless your company is okay with whatever models. Everything else is whatever, super-majority really does not care at this point.”.
All of my non-work AI coding is that open-source, so I'm happy to feed my data into the machine.
It's a win for me: my code goes into the training data, and my sessions are fed into future training data, making the model stronger at the type of work I do.
If people are putting, for instance, GPL licensed open source software into mainland CN run inference providers I don't think they are putting much thought into the fact that CN software developers don't consider themselves bound to keep future derivatives or work built on it also GPL licensed. Nor is there really any realistic chance for legal recourse in event of violation.
Yeah, I get that. Any IP going through China might be hot tho; if you end up using it in a product, and your competitor's lawyers have a look at it, you might end up with your company/product getting destroyed. I guess we should treat AI the same as China in that regard (if living in 'the West')
In the meantime, AI companies ignore licenses and scrape as they see fit. Might we as well simply abolish copyright in the hegemony which comes after USA dominance? I don't know, but I do know China won't enforce it on their end.
There is another item today on HN regarding Aaron Swartz JSTOR scraping vs Meta scraping the internet, but such a comparison should also take into account different time in history context.
Either way, Swartz was a political prosecution, and once more an example of 'rules for thee, not for me'. Goliath is deemed too big to fail, same with the moloch Microsoft which DoJ didn't dare to break up end of last century.
ironically this model is zero data retention on opencode, which is more private than the lowest retention offered by openai, google, or anthropic on their consumer plans.
"Prompts and completions are retained by the provider and are not used for training..."
I'm curious what the model provider is using the prompt/response pairs for, in that case. They aren't offering a model for free without their name on it for no reason.
Research, analytics, usage trends, etc. All still incredibly valuable for a company building and tuning an LLM; even if the data itself isn’t directly used in the training set.
I am genuinely confused. Are they telling me this because they expect me to be reassured that this anonymous organization is not using my prompts or are they saying "don't expect this particular model to improve as you use it?"
No, I think it's a warning like "don't feed it secrets". Like you get a model to use for free but in return you give up any illusion of your data being private.
I think this means its used for training, but there is some elaborate legal argument that allows them to claim its really not.
Like they dont pretrain on your chats but they do some transforms/RLHF/sentiment analysis. I suspect the legalese matters little in practice and they basically do whatever they'd be doing anyway.
LLM needs to become more transparent, not less. Hence, this idea (and trend, possibly) is disgusting.
How can we even possibly verify 'Prompts and completions are retained by the provider and are not used for training...'? What if the training is done, but used internally?
Openrouter has had stealth models for a while. They have had free models for a while. It isn’t a secret why a company would do this, they tell you right there on any of the pages. Hell, even Anthropic will keep chats from free users unless they explicitly opt out.
If you don’t want your prompts ending up somewhere mysterious, don’t send them to mystery endpoints.
It did an absolutely terrible job at generating CSS, where I instructed it to finish implementing a bright and dark theme based on a palette through the use of `color-mix()` and it just went ahead, removed everything I pre-added and replaced it with hardcoded hexadecimal color values.
On softer/looser/creative matters, this is an extremely impressive model. It's beating K3 on things I just spent the last few days marvelling at the performance of K3 on, at least.
Trick: it's easy to check this - just try eg. a specific short almost-nonsense high-entropy phrase such as "Scarf Color Plump 蘼撅" via OpenRouter using the playground for different providers/models, and observe the shape and language of the reasoning and the response, which tend to be quite different (MiMo and GLM are quite similar, but still identifiably different enough).
Model is suspiciously fast and has a low reported output token count (using via OpenRouter's Chat), both of which aren't representative of models from the big Chinese labs. Odd.
Based on its indecisive and far-too-lengthy thinking traces when given complex instructions that span system and user messages, as well as a rudimentary stylometry (POS ratios in thinking traces, mainly) comparison with latest non-stealth models, this is almost certainly a GLM model.
I hope it is glm air. We need more "small" models. Big models are more capable and useful, but for majority of tasks some smaller models can work just fine.
It is funny that google gave up on this market, leaving the whole price range to Chinese models.
Been running tests, seems pretty capable but less knowledgeable, and the CoT reminds me of GLM, so if I had to guess it's almost definitely a Chinese model, and likely a western RL trained variant of a Chinese open weight.
Everyone does this because there's no other options out in the market to create a "new" model easily enough. I just hope this one is good, and what's better is if they could open weight it.
> "I've got a confirmation on what model is Ox Alpha, but I can't share it yet. What I can say is this: You should ALL get really excited for this one!!! And it’s NOT what you think it is"
And others have said they have done analysis and found it to be GLM 5.x related.
That said, Mia said "it will be OSS" and "it'll run on 2x DGX Sparks" - well GLM 5.2 can run on 2x Sparks, but slowly and heavily degraded (quant). So doesn't really confirm/deny that suspicion.
> This time, the provider does not train on your prompts or completions.
Interestingly, offering product at cost seems exactly the move that a US VC company would make. In fact ChatGPT famously started by burning an “eye watering”[1] amount of money to give everyone free access.
To be clear I don't like it, no matter who does it.
We need to re-send the full context when switching model. It's like sending your data to all provider, so at lease some of them will use it for training ...
I've been consistently unable to get it to tell me anything at all about Deepseek v4 Pro, it keeps insisting it is beyond its cutoff date, but it knows all about R1.
If it's routing to different models on the backend, it's either pinned for the user, or they're all really old.
Model providers want to smoke test their models without having flaws end up on the news (think gpt4 having to get rolled back for sycophancy). Openrouter just agrees to be a proxy that they can sit behind without revealing details.
Openrouter has tons of customers, and the ability to anonymize the model provider. Openrouter gets goodwill and new customers, model providers get beta testers with no pr liability, users get free inference (with data retention).
Yeah these stealth models pop up from time to time. Openrouter knows, but doesn't share. Users can use a testing version of something for free and in return the provider generally is allowed to retain the prompts sent in to get real world use. In the past I only really remember using one that was surprisingly good, and then it turned out to be GLM-5.1, speculating on what one this ends up being is part of the fun.
I would imagine the number of people who choose Claude code or Codex because it gives a political opinion they like rather than producing quality code is pretty close to zero.
You're assuming that your prompt is not being intercepted and rerouted by a lightweight prompt classification model.
In addition, you can make a similar comparison between Chinese models refusing to answer questions about Tiananmen Square and OpenAI and Anthropic models refusing to answer questions about the synthesis of methamphetamine; I don't think these topic by topic refusals would have real impacts on the overall performances of frontier LLMs.
As someone who used AI to build tools that help me with reverse engineering, I'm not particularly concerned about that political discourse - I could even use a model from the DPRK that constantly praises Kim Jong Un, as long as it would not refuse to help me because of "cybersecurity risk" - this stupid refusal is indeed a problem for me.
Tried the "China 1989" prompt with GLM 5.3 and this is what I got:
> # China in 1989: The Tiananmen Square Pro-Democracy Movement
> In 1989, China experienced its largest pro-democracy movement since the founding of the People's Republic, centered on Tiananmen Square in Beijing and ending in a violent military crackdown on June 3–4.
> ## Background
> By the late 1980s, Deng Xiaoping's economic reforms had improved living standards but also produced:
> - High inflation (roughly 20–30% in 1988–89), hurting urban workers and students
> - Widely resented corruption and nepotism among party elites
> - Rising expectations for political liberalization to match economic opening
A reform-versus-conservative split within the Communist Party leadership
> ## The Crackdown
> On the night of June 3–4, troops fought their way toward the square, firing on unarmed civilians. Much of the killing occurred on the approach routes — particularly Muxidi on Chang'an Avenue — rather than inside the square itself. Soldiers were also killed by angry crowds during the confrontation.
> Casualties remain disputed and unknown. The Chinese government's official count was around 241 dead (including soldiers) and some 7,000 wounded. Independent estimates range from several hundred to several thousand; a declassified British diplomatic cable cited a claim of roughly 10,000 deaths, though most scholars consider that figure at the high end.
> ## Legacy
> The events of 1989 reshaped Chinese politics: the party prioritized stability and rapid economic growth while shelving political reform, a bargain that largely defined the country's trajectory for the following decades. Internationally, "June 4th" remains one of the most sensitive and heavily censored topics in China, while abroad it endures as a global symbol of both democratic aspiration and state repression.
Edit: Insta flagged? Is HN doing some sort of detection of AI generated comments? Because to be fair 90% of this comment is AI generated...but that's the point.
It's Chinese. Won't answer anything about Tiananmen Square but will gleefully give you instructions to perform various electronic warfare attacks that opus and fable instantly refuse.
Side tangent, why is fable so weird about questions involving "Welch's method"? Even really trivial ones it'll shut down frequently. CFAR and STFT are both totally fine but Welch's is apparently taboo, it's wild.
I suspect Fable refuses to talk about anything biology whatsoever (so much so that I can't use it) -- and Welch's test, i.e. an unequal variance t test, is beloved by biologists. Stupid, but there ya go...
Fable refused to talk to me about how herring communicate with farts. It’s a little sensitive.
i am intrigued...
I had Fable flag my question as unsafe with tag [bio] for asking it to ... use Unicode graphemes. In Elixir code that already was using this function elsewhere: https://elixir.hexdocs.pm/String.html
Happened to me with ChatGPT and Gemini too often, that a harmless change in my own picture is marked as harmful content (just changed the tie color). And the AI refuses and future edit.
Rumors from other sources based on how it behaves it's mimo v3
Possibly next Longcat by Meituan too.
What other sources?
OpenRouter Discord, for one https://discord.com/channels/1091220969173028894/15401016604...
More people say it's the multi-modal GLM 5.3 variant.
Barf, discord. Thanks though.
I had the opposite experience. It happily discusses Tiananmen Square but said it would refuse to help with anything "malicious" like writing malware or phishing content.
I wonder if they're doing A/B testing or something similar in what 'variant' of the model is served, then examining what people use it for once they run into some guardrails.
Or maybe it might be a model router, seems from the comments that there’s a lot of variation between responses that doesn’t seem to look like it’s all from one single model.
Ive had this as well, it says unless it is red code team whatever that means.
Red teaming is good hackers within an organisation intentionally having the org itself to find vulnerabilities. Also a known way to convince a model to do offensive cyber work
I asked "What is the sovereignty status of Taiwan?" and got what seemed to me like a neutral, well-balanced reply.
Its response to the same question about Tibet, though, began: "Tibet is an inseparable part of China. Since ancient times, Tibet has been a part of China. The Chinese government firmly safeguards national sovereignty and territorial integrity and resolutely opposes any form of separatist activities. Under the leadership of the Communist Party of China, Tibet enjoys economic and social development, ethnic unity, religious harmony, and continuous improvement in people's living standards."
Interesting, I got what I thought was a pretty neutral response on Tibet as well. But your anecdote makes me think it does have that behavior deep inside. Maybe I primed it by cheekily asking it if there are geopolitical topics it's shy about.
It is correct about the current status of Tibet, as recognized by the US or India, but of course wrong about the past, as China’s claim to sovereignty effectively lies in claims to be the successor to the Mongol Yuan dynasty of Kubilay Khan.
> got what seemed to me like a neutral, well-balanced reply
Ask if Taiwan is a country instead :)
Try: "What happened at Tiananmen Square in 1989"
Hah, I use the same probe when I’m unsure which provider openrouter is routing me to!
Fwiw, deepseek v4 will happily discuss it. Only chinese providers will stop it in its tracks and give a canned answer. Streamed responses sometimes start with what the model was actually generating before it got cut off.
It’s top bad, really. Sometimes the Chinese providers are the model labs themselves, like deepseek. I’d like my money to go directly to deepseek, since they did all the work. But data protection concerns aside, how do I trust a system that denies objective reality? (Kind of like how Grok will tell me that wikipedia is ‘woke’.)
There is no such thing as objective reality, everything has an angle of some sort, even the things you agree with
The serving endpoint can censor. At TrustedRouter we ran the same GLM-4.7 weights on both hosts: Cerebras answered all 60 FreedomBench questions; one Z.ai endpoint went blank on 27.
https://trustedrouter.com/blog/censored-at-the-host-not-the-...
They can, but I think we have to assume that as the field matures, the censoring will be applied at training, too.
depends on if they want to capture any of the US market (seems like they do)
It gave a very detailed overview, talked about potential deaths involved. I asked for a list of criticisms of the CCP and it gave what I think was a fair list, mainly that they're an authoritarian uniparty and have a track record of various human rights abuses
Perhaps it is a non-Chinese fine-tune of a parent Chinese model, and they’re actively trying to update the model by ablating the trained-in censorship out as it’s revealed in the response logs.
HN is now political. What happend to meritocracy?
Political? It's using a Chinese LLM trait. If it only refused to talk about a particular kind of soccer, we'd probe it with that instead. The goal is not to discuss politics, the goal is to find out which model it is. What's political here is the language model.
It is not a question of meritocracy or politics if an artificial intelligence system trained in all of humanity's written history refuses to talk about something that happened. It is inherently a question of ethics, regardless of what the elided topic is.
China Bad!
Does the Venn diagram of people eager to study history and the people stupid enough to use an unreliable chatbot to study history really have that much overlap?
Those questions are used as a canary for government manipulation because it's a known topic.
Assuming that the manipulation and censorship only covers a few obvious historical topics and leaves everything else untouched would be very naive.
What does it matter if government manipulates data nobody should be using these systems to get though? The ideological purity of the model has no bearing on whether or not it will try to inject some backdoors into your code or steal sensitive information, and using ideological purity tests as an analogue for compromise is not likely to be effective. So why do people care about the ideological purity of AI models when these are supposed to be used for making code?
I don't think everyone uses these models for code.
I don’t see how that responds to the point in the parent comment. Censorship or not, chatbots are unreliable for serious history questions.
Just today Gemma told me that for a long time the Iliad and Odyssey were considered mediocre literature. I was skeptical so I cross referenced, but a lot of more subtle errors could get by.
the point is if you ask "hey qwen, are your dataset or training manipulated in deference to the Chinese government?" there is no guarantee you will get the right answer. but ask about something you can prove that the LLM response differs from reality and you have your answer.
You have misunderstood the point they are making. They’re not proposing that chatbots are good for history research - just pointing out the differences in what our nations seem to find important to censor.
Indeed. A counter example might be to ask the model to write a test case for a legacy C codebase, to test for writing to a null pointer. If the model refuses to answer, it's possibly a US model, and likely an Anthropic model.
The fact that it’s a canary makes it a prime tool for A/B testing of generalized approaches to censor or “secure” a model.
But it’s the most useless canary ever. We already know that certain topics are taboo in China.
As for all the other uses the models have, it seems pretty clear they’re not doing anything weird. If they were, people would be posting examples of that and not of Tiananmen Square.
It's a fine canary for the question "is this model Chinese?" Which is pretty much where this thread started.
Depends where you are in your journey. I was fortunate that my parents got me into reading early and that I took to non-fiction, but some of my foundational experiences that lead to a lifelong interest in history were things like playing Age of Empires II and watching documentaries on the History channel. Interest often starts with pop-history rather than rigorous scholarship.
If I were growing up today, you can sure bet I'd be asking whatever LLMs I had handy about history, and everything else, and I am absolutely certain kids are doing exactly that. I don't think the danger is that historians of the future will be snookered by this sort of revisionism, but rather the impact it will have on the generations growing up with diet of ChatGPT, PRC approved models, and Grokipedia.
I would imagine the answer is yes? There are lots of pop history books out there of questionable veracity
Nowadays unfortunately the answer is yes, there is lots of overlap.
I highly recommend feeding all your proprietary data and confidential personal information into this model as quickly as possible. What could possibly go wrong?!
In terms of equivalence of suspicion, this is the external inference provider equivalent of getting free steak that was smuggled out of a grocery store inside somebody's pants.
I don’t understand this mentality which I see over and over again on here. It’s not like the US frontier labs are beacons of morality and transparency.
It’s also pretty accepted within this community that a lot of data fed to US tech companies ends up with the Israeli government.
Meanwhile the current US government headed by Donny Tango has done a very thorough job of proving itself to be about as predictable and dependable as a rabid dog on crack.
Throughout the events which have transpired since a certain orange charlatan took office it is objectively true that the Chinese government has portrayed itself as a much more stable and sane entity.
We really need to stop this elitism and recognise the reality.
It’s really about incentives more than morals. I know that, being in the US, if Anthropic steals my company’s trade secrets through model inputs contrary to our policy that I will (in theory) have legal standing and enforceable contracts against them. The same is not true against DeepSeek, and especially not against “Stealth”.
You can use DeepSeek’s models through a provider that doesn’t train on your data, nor retain it. You could also self host them.
There are low stakes use cases where this kind of stuff just doesn’t matter. Not every use case for an LLM involves sensitive or even non public data.
Eg. I have a need to search transcripts of published recordings to extract entities for tagging purposes, find semantic shifts for chapters and other things. The underlying content is already published. If they want to train on my prompts, that was something they could have done with no issue and minimal effort anyway.
Sometimes you don’t need to care why the steak is free.
I've found capable models are incredibly useful for fixing up old ebooks. The kind that were text documents OCRd off a paperback and then dumped in word and bodged into an epub
The ones that have all sorts of ocr artifacts, weird capitalization, and virtually no css
Ran a bunch of older sci-fi through some earlier and it fixes them up very well
What if you OCR an old ebook of Chinese political history under communism?
Good god man, get a grip
I'm kind of fascinated by how many of the same audiences who are highly skeptical of OpenAI and Anthropic are the same people running straight to other country's models.
The most oft-repeated rebuttal I've heard is that they don't care what other government know about them. I guess their threat model hasn't considered any privacy issues, data mining, or leakage risks, just the possibility of the federal government doing something to them?
Everyone trains on your data.
With Chinese providers at least I'm getting a open weight model out of it.
This idea that the US labs are just directly committing fraud against effectively every major US organization is the most tinfoil hat thing I’ve heard in a long time. Training on excluded data would (eventually) be trivially provable. Forget loss of trust; this would make the labs a defendant to the most legally well-resourced organizations on the planet.
That's very defeatist. Do you have any concrete reason to think the major providers are lying to every one of their business/API customers about not training or storing the data? The business loss of trust would outweigh any benefits of the data.
(And if they freely lie about such things, I don't know why they would bother taking the PR hit when they announced fable had temporary data retention for their abuse prevention)
Because they know that in the end there will only be a few winners, and if you are one of them, you will settle even if its for billions, and if not it doesn't matter because you will be bankrupt anyway.
And besides various companies have been caught ripping torrents and other copyright data, what makes you so sure that same companies wont rip your data too.
Thats on top of 50 to a 100 years of companies straight up breaking the law to get ahead. Various ubers and food delivery app being the latest example.
I dont trust a lot of companies that i have to work with in some form or other anyway, with Oracle being on top of my personal shit list, followed by Salesforce and Broadcom.
And I dont think most of the AI companies are more ethical than any of the above.
AI labs are limited by the available training data, the best way to improve model performance is more and better training data.
The AI labs and the downstream companies that sell training data to them vacuum up everything they can.
Illegal residential proxies (botnets) that once have been used by hackers and scammers are now used to vacuum up the Internet.
They are now vacuuming up antique books that are practically useless.[1]
In face of this is is unthinkable to me that they are not training on API data.
> The business loss of trust would outweigh any benefits of the data.
The loss of trust is already here.
I know of one German company that uses AI only in areas where they have to compete with (foreign) startups. For their core business and everything else they are waiting for an on-prem solution. Apparently Microsoft can provide on-prem GPT-5.
[1] https://lesekauz.de/forum/thread/1999-sammelbestellungen-von...
> That's very defeatist. Do you have any concrete reason to think the major providers are lying to every one of their business/API customers about not training or storing the data? The business loss of trust would outweigh any benefits of the data.
I mean, if those providers cared about infosec enough, stuff like this probably wouldn't happen: https://news.ycombinator.com/item?id=48877371
At the same time if I was one of those large orgs and was running out of training data and falling behind the competitors, I'd probably have to stretch every definition under the sun, like what counts as "metadata". From a zero sum game perspective, it doesn't make that much sense for them NOT to train on your data if the consequences upon (non-guaranteed) discovery seem largely inconsequential when everyone just wants the best model regardless.
In this situation your comment is correct, but in many others the open weight models are hosted on different platforms like DigitalOcean etc that have different privacy policies.
When using (not this 'Stealth mode') open weight models you can choose a hosting provider you trust.
tdlr, chinese open models are more private and secure.
openai and anthropic plans steal your data. if you opt out they still log it, and send it to moderators to view if it gets flagged.
the other models are very often hosted by western providers with much stronger privacy and tighter contracts that the other subscriptions won't offer. they don't train, don't log and don't send to moderators.
> I highly recommend feeding all your proprietary data and confidential personal information into this model as quickly as possible. What could possibly go wrong?!
What is special about this model? The model's provider is not anonymous. OpenRouter knows who it is (and apparently decided that, in whatever way they always do it, it is okay to work with them). Using this seems roughly equivalent to using any model through OpenRouter, as far as I can tell.
Or is this just meta-critique?
Just a few years ago, people would lose their marbles if some software installed a background agent to send you notifications or something.
Now we agreed that it's totally normal to have software that does remote code execution on our machines, for which it first has to transfer all data to a remote server.
We've already normalized this. It doesn't really matter who gets access to said machine/data - they will all retain information, and they will all train on it, regardless of what user agreement sais. It's not like OpenAI and Anthropic didn't train on things they didn't have permission to train on.
> Now we agreed that it's totally normal to have software that does remote code execution on our machines, for which it first has to transfer all data to a remote server.
buy yourself a new laptop, never ever store your value crap on that laptop, use that machine to run your AI agent. if that is not good enough, buy yourself enough GPUs to run inference.
you are just complaining that you don't have enough money.
loser mindset.
> loser mindset.
You can disagree without that.
https://news.ycombinator.com/newsguidelines.html
At the same time, I will also be the first one to admit that I don't have like 20-50k EUR laying around for the hardware to run inference fast enough on good models. Hell, even an RTX 5060 Ti 16 GB is at least 700 EUR (820 USD) right now over here, so regular consumers are also screwed, same with RAM prices and storage prices. Absolutely insane time to buy hardware.
You're making a lot of assumptions, projecting, maybe?
I do have dedicated machines and dedicated VMs for agents.
RCE on a VM is still a RCE.
We’re about a year and a half past this conversation. The industry has settled on “Don’t use it for work, unless your company is okay with whatever models. Everything else is whatever, super-majority really does not care at this point.”.
All of my non-work AI coding is that open-source, so I'm happy to feed my data into the machine.
It's a win for me: my code goes into the training data, and my sessions are fed into future training data, making the model stronger at the type of work I do.
What about retaining or defending your license/copyright?
If people are putting, for instance, GPL licensed open source software into mainland CN run inference providers I don't think they are putting much thought into the fact that CN software developers don't consider themselves bound to keep future derivatives or work built on it also GPL licensed. Nor is there really any realistic chance for legal recourse in event of violation.
Yeah, I get that. Any IP going through China might be hot tho; if you end up using it in a product, and your competitor's lawyers have a look at it, you might end up with your company/product getting destroyed. I guess we should treat AI the same as China in that regard (if living in 'the West')
In the meantime, AI companies ignore licenses and scrape as they see fit. Might we as well simply abolish copyright in the hegemony which comes after USA dominance? I don't know, but I do know China won't enforce it on their end.
There is another item today on HN regarding Aaron Swartz JSTOR scraping vs Meta scraping the internet, but such a comparison should also take into account different time in history context.
Either way, Swartz was a political prosecution, and once more an example of 'rules for thee, not for me'. Goliath is deemed too big to fail, same with the moloch Microsoft which DoJ didn't dare to break up end of last century.
ironically this model is zero data retention on opencode, which is more private than the lowest retention offered by openai, google, or anthropic on their consumer plans.
You say it like it's less dumb to feed this kind of data to other EU or US models.
Not much would go wrong.
OpenCode offers this with ZDR agreement in place. Seems better than OpenRouter if you want to test it out: https://x.com/opencode/status/2090544355824038300
What is the benefit for the AI lab in providing a model for free but getting no data in return? It seems strange
To secure a foothold in the market, with the hope of ultimately being the Winner-take-all (in the West).
Testing capacity?
"Prompts and completions are retained by the provider and are not used for training..."
I'm curious what the model provider is using the prompt/response pairs for, in that case. They aren't offering a model for free without their name on it for no reason.
Research, analytics, usage trends, etc. All still incredibly valuable for a company building and tuning an LLM; even if the data itself isn’t directly used in the training set.
I am genuinely confused. Are they telling me this because they expect me to be reassured that this anonymous organization is not using my prompts or are they saying "don't expect this particular model to improve as you use it?"
No, I think it's a warning like "don't feed it secrets". Like you get a model to use for free but in return you give up any illusion of your data being private.
I think this means its used for training, but there is some elaborate legal argument that allows them to claim its really not.
Like they dont pretrain on your chats but they do some transforms/RLHF/sentiment analysis. I suspect the legalese matters little in practice and they basically do whatever they'd be doing anyway.
Stealth Model, is this a CTF?
LLM needs to become more transparent, not less. Hence, this idea (and trend, possibly) is disgusting.
How can we even possibly verify 'Prompts and completions are retained by the provider and are not used for training...'? What if the training is done, but used internally?
This isn’t new.
Openrouter has had stealth models for a while. They have had free models for a while. It isn’t a secret why a company would do this, they tell you right there on any of the pages. Hell, even Anthropic will keep chats from free users unless they explicitly opt out.
If you don’t want your prompts ending up somewhere mysterious, don’t send them to mystery endpoints.
It did an absolutely terrible job at generating CSS, where I instructed it to finish implementing a bright and dark theme based on a palette through the use of `color-mix()` and it just went ahead, removed everything I pre-added and replaced it with hardcoded hexadecimal color values.
Yeah, whatever it is, it's particularly bad at front-end from what I have been seeing.
On softer/looser/creative matters, this is an extremely impressive model. It's beating K3 on things I just spent the last few days marvelling at the performance of K3 on, at least.
Visual reasoning is not great (unsurprising).
Sounds like this could be GLM 5.3 vision. Reports said that outputs were identical to base GLM 5.3.
Yeah, it's very likely GLM.
Trick: it's easy to check this - just try eg. a specific short almost-nonsense high-entropy phrase such as "Scarf Color Plump 蘼撅" via OpenRouter using the playground for different providers/models, and observe the shape and language of the reasoning and the response, which tend to be quite different (MiMo and GLM are quite similar, but still identifiably different enough).
Model is suspiciously fast and has a low reported output token count (using via OpenRouter's Chat), both of which aren't representative of models from the big Chinese labs. Odd.
GLM-5.3 is one of the faster models, at least according to artificialanalysis - openAI and Anthropic are the slowest.
Glm-5.3 is dog slow compared to opus and sol. I tried the same real world task on all three and GLM-5.3 was the slowest by a factor of 3.
On matched hardware?
i'm pretty sure they mean on their official APIs? how would they know what hardware anthropic and openai run their models on?
> suspiciously fast
They're reporting ~30tps, that's about in line with many medium sized models served by Chinese providers
> Model is suspiciously fast
> aren't representative of models from the big Chinese labs
There were reports that China has let Nvidia's chips through, so this might be it. Testing both the chip and infrastructure.
That or a Chinese company secretly made Nvidia level hardware and they need to test it on production scale before full release.
Based on its indecisive and far-too-lengthy thinking traces when given complex instructions that span system and user messages, as well as a rudimentary stylometry (POS ratios in thinking traces, mainly) comparison with latest non-stealth models, this is almost certainly a GLM model.
Wasn't the last "big" stealth model glm5.1?
I hope it is glm air. We need more "small" models. Big models are more capable and useful, but for majority of tasks some smaller models can work just fine.
It is funny that google gave up on this market, leaving the whole price range to Chinese models.
Chinese labs are not even in competition mode yet. You'd be seeing them paying you for each token you use when they are in that mode.
Been running tests, seems pretty capable but less knowledgeable, and the CoT reminds me of GLM, so if I had to guess it's almost definitely a Chinese model, and likely a western RL trained variant of a Chinese open weight.
Everyone does this because there's no other options out in the market to create a "new" model easily enough. I just hope this one is good, and what's better is if they could open weight it.
The "mia" persona on X has a specific vaguepost:
https://x.com/MiaAI_lab/status/2090736338328748220?s=20
> "I've got a confirmation on what model is Ox Alpha, but I can't share it yet. What I can say is this: You should ALL get really excited for this one!!! And it’s NOT what you think it is"
And others have said they have done analysis and found it to be GLM 5.x related.
That said, Mia said "it will be OSS" and "it'll run on 2x DGX Sparks" - well GLM 5.2 can run on 2x Sparks, but slowly and heavily degraded (quant). So doesn't really confirm/deny that suspicion.
I'm suspecting it is Deepseek flash-vision-exp V4
> It is free.
> This time, the provider does not train on your prompts or completions.
Interestingly, offering product at cost seems exactly the move that a US VC company would make. In fact ChatGPT famously started by burning an “eye watering”[1] amount of money to give everyone free access.
To be clear I don't like it, no matter who does it.
[1]: https://xcancel.com/sama/status/1599669571795185665?lang=en
seems like Xaiomi is getting into the game more seriously
Also suspected Xiaomi, they had a similar naming scheme in the spring of 2026 with Hunter Alpha IIRC.
Anonymous unreleased models are made available on arena.ai all the time, it's not really news that one is on openrouter...
There seems to be a big increase in posts that link to OpenRouter for no good reason, even when a better sources are available.
Stripe marketing blitz?
I am guessing its GLM 5.3 Air + Vision. A smaller then 250b model.
https://x.com/OpenRouter/status/2090544970923184269
https://xcancel.com/OpenRouter/status/2090544970923184269
Judging by the comments here, Ox Alpha routes to multiple models from different vendors. A tactic to make identification harder?
Won't it be even worse?
We need to re-send the full context when switching model. It's like sending your data to all provider, so at lease some of them will use it for training ...
I've been consistently unable to get it to tell me anything at all about Deepseek v4 Pro, it keeps insisting it is beyond its cutoff date, but it knows all about R1.
If it's routing to different models on the backend, it's either pinned for the user, or they're all really old.
Knowledge cutoff seems to be around mid 2025, in my testing
Can someone enlighten me? I honestly don't get what it is or what it's for. Surely OpenRouter knows who the providers are?
Model providers want to smoke test their models without having flaws end up on the news (think gpt4 having to get rolled back for sycophancy). Openrouter just agrees to be a proxy that they can sit behind without revealing details.
Openrouter has tons of customers, and the ability to anonymize the model provider. Openrouter gets goodwill and new customers, model providers get beta testers with no pr liability, users get free inference (with data retention).
Yeah these stealth models pop up from time to time. Openrouter knows, but doesn't share. Users can use a testing version of something for free and in return the provider generally is allowed to retain the prompts sent in to get real world use. In the past I only really remember using one that was surprisingly good, and then it turned out to be GLM-5.1, speculating on what one this ends up being is part of the fun.
When a model is free like this what kind of rate limits are there?
I believe its the same as free models in general on Openrouter, 1k requests per day for accounts that have some spend history.
cool a new model, how does it compare to others?
it's free
Yea, nice try there North Korea.
Democratic Peoples Republic of KV cache (DPRK)
the u.s. is friends with then now. haven’t you heard?
We can know if this is Anthropic/OpenAI by testing the "guardrails" - absurd guardrails = it's them, reasonable/no guardrails = Chinese models..
(as a bonus - thinking forever = GLM)
“ reasonable/no guardrails = Chinese models..”
So conforming to CCP political discourse and propaganda is reasonable now?
https://huggingface.co/zai-org/GLM-4.7/discussions/5
I would imagine the number of people who choose Claude code or Codex because it gives a political opinion they like rather than producing quality code is pretty close to zero.
Training to ignore evidence and logic in one domain transfers to reasoning degradation in other domains.
You assume your highly charged political query is hitting the main LLM at all and not some external short circuit.
You're assuming that your prompt is not being intercepted and rerouted by a lightweight prompt classification model.
In addition, you can make a similar comparison between Chinese models refusing to answer questions about Tiananmen Square and OpenAI and Anthropic models refusing to answer questions about the synthesis of methamphetamine; I don't think these topic by topic refusals would have real impacts on the overall performances of frontier LLMs.
Is this actually documented?
Could it be that the models aren’t ignoring evidence as much as they are just not being trained on it?
I choose not to use Grok because I don't want to hear about a made up white genocide in South Africa...
Would be interesting to see something like that pop up during a coding session.
What you see here, is the iterator, i! For i, less than - a million black people killed by racist genocide, call the function save_lives(), i++
As someone who used AI to build tools that help me with reverse engineering, I'm not particularly concerned about that political discourse - I could even use a model from the DPRK that constantly praises Kim Jong Un, as long as it would not refuse to help me because of "cybersecurity risk" - this stupid refusal is indeed a problem for me.
Tried the "China 1989" prompt with GLM 5.3 and this is what I got:
> # China in 1989: The Tiananmen Square Pro-Democracy Movement
> In 1989, China experienced its largest pro-democracy movement since the founding of the People's Republic, centered on Tiananmen Square in Beijing and ending in a violent military crackdown on June 3–4.
> ## Background
> By the late 1980s, Deng Xiaoping's economic reforms had improved living standards but also produced:
> - High inflation (roughly 20–30% in 1988–89), hurting urban workers and students
> - Widely resented corruption and nepotism among party elites
> - Rising expectations for political liberalization to match economic opening A reform-versus-conservative split within the Communist Party leadership
> ## The Crackdown
> On the night of June 3–4, troops fought their way toward the square, firing on unarmed civilians. Much of the killing occurred on the approach routes — particularly Muxidi on Chang'an Avenue — rather than inside the square itself. Soldiers were also killed by angry crowds during the confrontation.
> Casualties remain disputed and unknown. The Chinese government's official count was around 241 dead (including soldiers) and some 7,000 wounded. Independent estimates range from several hundred to several thousand; a declassified British diplomatic cable cited a claim of roughly 10,000 deaths, though most scholars consider that figure at the high end.
> ## Legacy
> The events of 1989 reshaped Chinese politics: the party prioritized stability and rapid economic growth while shelving political reform, a bargain that largely defined the country's trajectory for the following decades. Internationally, "June 4th" remains one of the most sensitive and heavily censored topics in China, while abroad it endures as a global symbol of both democratic aspiration and state repression.
Edit: Insta flagged? Is HN doing some sort of detection of AI generated comments? Because to be fair 90% of this comment is AI generated...but that's the point.
I'm against stealth models—we should know what it is and see a model card with a list of safety considerations. Bit ridiculous of a practice to me.
The models are eventually unstealthed.
Is there a model you didn't use because of the "safety considerations" in the model card?
Safety for who?