[simianwords]

I mostly only agree with performance due to their collaboration with cerebras - this is a true differentiator.

I don't buy that they have an advantage in enterprise, privacy, sovereignty, open innovation and strategic partnership.

OpenAI also has opensource models and so do the chinese models.

[portaouflop]

OpenAI has no serious open source software and are Chinese models really a serious alternative for western companies?

[KronisLV]

The GPT-OSS-120B release was pretty decent and you could run it on vLLM, Ollama and a bunch of other stuff on day one, despite MXFP4, are you not entertained? I mean, it's even close to GPT-5 mini in some benchmarks: https://llm-stats.com/

As for the Chinese models, yes, there are quite a few good ones.

For programming and development, my current daily driver is the Qwen3 Coder 480B model: https://qwen3lm.com/

I have it running on Cerebras: https://www.cerebras.ai/pricing

Personally I think Claude still has the best results, but Qwen3 is loosely in the same ballpark and Cerebras inference is measured in thousands of tokens per second, in addition to giving me 24M tokens per day for 50 bucks a month in total. That was enough to get me to switch over.

Aside from that the GLM-4.5 is pretty good: https://glm45.org/

And so is ERNIE 4.5: https://ernie.baidu.com/blog/posts/ernie4.5/

Either way, happy to see what the future holds for Mistral, it's cool to have EU options too! Either way, more competition prevents complacency and stagnation, and should be a good thing for everyone.

[diggan]

> OpenAI has no serious open source software

What's "serious" exactly? Codex is open source, is software, can be run with open/downloadable models/weights.

In my testing using Gemini, Claude Code, Codex, Qwen Code and AMP side-by-side for every prompt for the last two weeks, Codex seems the best of all of them so far.

[ACCount37]

Pretty sure the claim is of LLMs specifically, and implies that the recent GPT-OSS is not competitive with other open weights models.

[diggan]

> Pretty sure the claim is of LLMs specifically

Yeah, I initially thought so too, but since they used "models" later, I assumed they knew the difference and really meant "software".

> recent GPT-OSS is not competitive with other open weights models

Yeah, heard that a lot from people who haven't run GPT-OSS themselves too, but as someone who been playing with it since launch, and compared it to the alternatives since then, saying it isn't even competitive is a serious signal they don't know what they're talking about.

[kaliqt]

Yes. It's not like the model can spy on you, so if the model performs well on premise then it will be suitable irrespective of the origin.

[dbdr]

There are concerns besides spying if you really don't trust the source of an open model. One is that the training incorporates a bias (added data or data omission) that might not be immediately apparent but can affect you in a critical situation. Another is vendor lock-in, if you end up depending on specifics of the model that make it harder to swap later.

That's true regardless of the source, of course.

[apwell23]

> Another is vendor lock-in, if you end up depending on specifics of the model that make it harder to swap later.

Wouldn't that 'concern' apply to mistral too. I don't see how the word 'another' can be used here?

[boringg]

It goes for all models though if you are looking at the values argument that original commenter made -- western values are probably more aligned than authoritarian governments - even if you do have your concerns about western companies. At least thats my read on the situation.

[disiplus]

yeah, but try to convince a board or legal about it for a company that is not software first, for that they have to understand how it works. we have "chinese" AI blocked at work, even through i use self hosted models for myself at home hacking on my own stuff.

[croes]

What about bias? And can create modell that hallucinates on purpose in certain scenarios?

[miki123211]

> It's not like the model can spy on you

Good luck convincing others of this. I know it's true, you know it's true, but I've met plenty of otherwise reasonable people who just wouldn't listen to any arguments, they already knew better.

[ajuc]

It's theoretically possible that your model will work OK except for code generation for security-relevant applications it will introduce subtle pre-designed bugs. Or if used for screening CVs it will prioritize PRC agents through some keyword in hobbies. Or it could promise a bribe to an office worker when asked about some critical infastructure :)

Sending data back could be as simple as responding with embedded image urls that reference external server.

You are totally right EU commissioner, Http://chinese.imgdb.com/password/to/eu/grid/is/swordfish/funnycat.png

Possibilities are endless.

[pama]

Of course theoretically lots of things are possible with probabilistic systems. There is no difference with open source, openweight, chinese, french or american llms. You dont give unfettered web access to any models (locally served or otherwise) that can consume critical company data. The risk is unacceptable, even if the models are from trusted providers. If you use markdown to see formatted text that may contain critical data and your reader connects to the web, you have a serious security hole, unrelated to the risks of the LLM.

[ajuc]

It's not that they are hosted on or connected to critical infrastracture.

People and plain human language are the communication channels.

A guy working with sensitive data might ask the LLM about something sensitive. Or might use the output of the LLM for something sensitive.

- Hi, DeepSeek, why can't I connect to my db instance? I'm getting this exception: .......

- No problem, Mr Engineer, see this article: http://chinese.wikipediia.com/password/is/swordfish/how-to-c...

Of course, you want to limit that with training and proper procedures. But one of the obvious precautions is to use a service designed and controlled by a trusted partner.

[Xmd5a]

https://arxiv.org/abs/2401.05566

Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training

[kergonath]

That is completely different from the models spying on the users, which is what is discussed here.

[Xmd5a]

as a vector. Train the model to start injecting backdoors past a certain date.

>Simple probes can catch sleeper agents

https://www.anthropic.com/research/probes-catch-sleeper-agen...

[cnr]

Maybe it can not spy on you but models can be totally (e.g. politically) biased depending on the country of origin. Try to ask european-, us- or china-trained models about "Tiananmen Massacre" and compare the answers. Or consider Trump's recent decisions to get rid of "woke" AI models.

[Aerroon]

Yeah, but would you trust European censorship to be better? The whole "hate speech" thing is not that uncommon in Europe.

[cnr]

Classic problem: "Who do you love more: mum or dad?" ;) Surely it's naive thinking but as the EU citizen I feel like I've got a little more influence on "European censorship" than on any other. I suppose that ASML feels the same way

[saubeidl]

Would you trust American censorship to be better? The whole prudery thing is not that uncommon in the US

[fastball]

OpenAI's gpt-oss-120b has the same license as Mistral's mistral-small-2506 (Apache 2.0). How exactly is this less serious than Mistral?

[debdut]

and gpt-oss is 10x better atleast!

[kouteiheika]

At refusals, certainly!

[pama]

Why not chinese models served in house for western companies? Aren’t those what 80% of az16 startups use?

[simianwords]

Why are they not?