Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Mistral built a $14B AI empire by not being American

forbes.com

Post reply on HN
Print view View original

151–160 of 187 posts

Re: Mistral built a $14B AI empire by not being American

#151

Post by badsectoracula »

0xbadcafebee wrote: post #84
phillc73 wrote: post #3

I am a Mistral Le Chat Pro subscriber. I specifically chose to test their offerings because they are European. I don't have the necessary local hardware to run really big models, therefore need to choose a cloud provider if I want LLM action. I find the antics of Anthropic, OpenAI, Google, Microsoft distasteful and avoid their products where I can. After testing Le Chat and Devstral-2 for a while, I felt their offeri…

Mistral models are definitely good enough. Most people fall for what I call the SOTA Logical Fallacy : whenever there is a 'better model', they think they need to use it, when less-powerful models actually perform the same tasks just as well. (it's an inverse form of the Shifting Baseline Syndrome : every time a new model comes out, people shift their baseline of what is acceptable, despite the fact that a previous b…

TBH sometimes i feel like i'm "emotionally attached" to Mistral's models because i always end up using them :-P. However that is because, as you wrote, their small models (i only use local stuff) are very strong. In fact i was trying Qwen3.6 27B recently and while it is nice that it can do tool calls during the reasoning process (i had it confirm its thoughts by writing Python code) it often ended up confusing itself (regardless of tool calls) during reasoning, ending up in loops where it questions itself over and over endlessly.

Devstral Small 2 however just works, for the most part. Qwen3.6 27B can probably handle more complex tasks (when i asked it as a test to write a function that checks for collision between two AABBs in C and gave it a tool to call Python code for confirmation, it actually wrote a Python script that writes C code with the tests, then calls GCC to compile the C code and runs the binary to run the tests, which is something Mistral's small models couldn't do) but i always felt i can just leave DS2 doing stuff in the background (or when i'm doing something else) and it'll produce something relatively useful whereas the little time i spent with Qwen3.6 27B it felt more "unstable" (and much slower, both because of literally slower inference and because of endless reams of text).

Recently i also started using Ministral 3B and 14B - these can do some reasoning too and for very simple stuff Ministral 3B is very fast (i actually didn't expect a 3B model to be anything more than novelty) and have some vision abilities (though they're quite mediocre at vision so i haven't found much use for this - passing something via GLM-OCR to extract all text and feed it to another model feels more practical).

Also as i wrote in another comment, every Mistral model i've tried never questioned me, which i certainly prefer

Re: Mistral built a $14B AI empire by not being American

#152

Post by cpburns2009 »

NekkoDroid wrote: post #138
mring33621 wrote: post #85

Earlier quoted context omitted.

Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf. This 7.5GB model runs well in llama.cpp on my 2021 Macbook Pro and is good at both coding and business document analysis tasks.

> Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf. Thiss sounds like such a shitpost I initially thought you were joking... but this seems to be a real model???

There's a method to the madness:

- Mistral-Nemo: the actual model developed by Mistral and Nvidia.

- 2407: likely the release date of the base model, July of 2024.

- 12B: the model has 12 billion parameters.

- Thinking: the model operates in thinking mode (generates output plan and injests it before producing actual output).

- Claude-Gemini-GPT5.2: I think this means the model was finetuned with session data from Claude, Gemini, and GTP5.2 to replicate their behavior.

- Uncensored-HERITIC: the model was uncensored using the automated Heretic method.

- Q4_k_m: the model is quantized (lossy compression) to ~5 bpw from orignal 16 bpw.

Re: Mistral built a $14B AI empire by not being American

#153

Post by amunozo »

0xbadcafebee wrote: post #84
phillc73 wrote: post #3

I am a Mistral Le Chat Pro subscriber. I specifically chose to test their offerings because they are European. I don't have the necessary local hardware to run really big models, therefore need to choose a cloud provider if I want LLM action. I find the antics of Anthropic, OpenAI, Google, Microsoft distasteful and avoid their products where I can. After testing Le Chat and Devstral-2 for a while, I felt their offeri…

Mistral models are definitely good enough. Most people fall for what I call the SOTA Logical Fallacy : whenever there is a 'better model', they think they need to use it, when less-powerful models actually perform the same tasks just as well. (it's an inverse form of the Shifting Baseline Syndrome : every time a new model comes out, people shift their baseline of what is acceptable, despite the fact that a previous b…

For certains tasks that are not hard but depend a clear specification, it's even better to haver less capable model because it forces you to do a better description of what you want, ending up with a better results. I will defend my PhD thesis soon and I will buy a yearly Mistral subscription at a student price to get it for cheap.

Re: Mistral built a $14B AI empire by not being American

#154

Post by NekkoDroid »

cpburns2009 wrote: post #152
NekkoDroid wrote: post #138

Earlier quoted context omitted.

> Try Mistral-Nemo-2407-12B-Thinking-Claude-Gemini-GPT5.2-Uncensored-HERETIC_Q4_k_m.gguf. Thiss sounds like such a shitpost I initially thought you were joking... but this seems to be a real model???

There's a method to the madness: - Mistral-Nemo: the actual model developed by Mistral and Nvidia. - 2407: likely the release date of the base model, July of 2024. - 12B: the model has 12 billion parameters. - Thinking: the model operates in thinking mode (generates output plan and injests it before producing actual output). - Claude-Gemini-GPT5.2: I think this means the model was finetuned with session data from Cla…

Yea, I know what the parts individually mean. I just meant as a whole it just seemed so obsurd.

Re: Mistral built a $14B AI empire by not being American

#155

Post by overfeed »

hsuduebc2 wrote: post #144
ta20240528 wrote: post #105

Earlier quoted context omitted.

You are incorrect: 1. the 2018 CLOUD Act mandates US companies — and their subsidiaries — to provide information to the US government on demand, regardless of where the data is stored 2. FISA secret courts prevent companies from even saying they where summoned, or telling anyone who or what the case was about (including canaries). So you won't ever know if your data was handed over to the US government.

They should be legally and physically separated and these actions should be then potentially illegal for Europeans so I do not think I'm at least infactual. But assuming the owner is US company abiding US laws it's safe to assume that data would be transferred to US one way or the another.

The US intelligence machinery spied on Angela Merkel's phone. Do you suppose secretly demanding cooperation for Lawful intercept capabilities in Amazon GmbH is somehow beyond or beneath them?

Also consider that all communication between the European subsidiaries to the HQ is fair game under FISA.

Re: Mistral built a $14B AI empire by not being American

#156

Post by overfeed »

cbg0 wrote: post #104
FpUser wrote: post #102

Earlier quoted context omitted.

Unless it is air gapped which it is not there is no way to protect Amazon's developed and owned software stack from reporting back to headquarters.

Sure there is: contracts, laws and prison time can ensure that doesn't happen.

The European leaders would have have no say in it. If the software from Seattle is designed to covertly exfiltrate information, they won't even know it. Even if they review the individual code changes, it can be an obfuscated attack similar to XZ where the code itself is clean, but not so much for the network fabric firmware binary test data.

Re: Mistral built a $14B AI empire by not being American

#157

Post by ben_w »

aurareturn wrote: post #74
ben_w wrote: post #31

Earlier quoted context omitted.

> Edit: I hear the commenters to this post. However, Mistral still relies on American chips. If there is truly a divorce between Europe and the US such that relying OpenAI or Anthropic is not an option, neither will relying on Nvidia and likely the thousands of smaller hardware and software suppliers that make Mistral work. That's why I don't think it's realistic to say that Europeans can't rely on OpenAI/Anthropic a…

American designed and controlled by the US government. See China export ban.

Indeed. But that US government control is limited to adding restrictions and removing them again, or offering money, it can't magic things out of the air when other people or nations put restrictions on them. And that's even absent Trump being an idiot who had to be talked out of killing the goose while it was laying golden eggs: https://finance.yahoo.com/news/trump-considered-breaking-nvi...

For example, if China looks at the chaos that has been Russia attempting to take Ukraine and the USA attempting to control Iran and thinks "Amateurs" right before doing the same to Taiwan, the GPU supply takes a massive dive. And if North Korea goes after South Korea, RAM gets even harder to buy.

And if the EU says no more ASML sales outside the EU, that delays factories outside the EU by a few more years than they'd otherwise take.

But in the other direction, if NVIDIA's only thing is IP, and the IP is tied to a nation which thinks everyone else on the planet is hostile, that IP may not get protected very well. Right now this is unthinkable, but 5 months ago so was Trump threatening force to take Greenland.

Re: Mistral built a $14B AI empire by not being American

#158

Post by tasuki »

badsectoracula wrote: post #147
tmikaeld wrote: post #139

Earlier quoted context omitted.

My biggest issue with Devstral and even their biggest model is that they’re dangerous unless closely directed and reviewed and i mean CLOSELY. Unfortunately mistral models will believe and do anything. See: https://petergpt.github.io/bullshit-benchmark/viewer/index.v... See some of the test results, it’s horrifying

FWIW personally i prefer this. When i tried Qwen3.6 and asked it a few questions, while it did respond, it was ADAMANT i should do something else when i really wanted an answer to the question i made. It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific que…

> It felt like when you search something and a stackoverflow answer about what you search for comes up and the most upvoted answer is about using/doing something else - when you want a specific answer to that specific question, not something else.

Don't you think there's usually a good reason for this? Whenever this happened to me, the problem was my ignorance.

Re: Mistral built a $14B AI empire by not being American

#159

Post by joe_mamba »

philipallstar wrote: post #135
troyvit wrote: post #126

Earlier quoted context omitted.

FTA: > So Mistral is developing its own data centers, starting with one outside Paris. Mensch projects it will have 200 megawatts of capacity by the end of 2027. Power from France’s state-owned nuclear plants will help, but the buildout could still cost an estimated $5 billion. Mensch tapped oil-rich Abu Dhabi and reportedly sought debt financing to help pay for it. Though to your point it won't be running until 2027…

Yes, I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool) for very long time. Even those data centres, while run by Europeans, are still being made with almost entirely US tech. But at least the EU companies can borrow some oil money and buy in the stuff developed by someone else's R&D spend, which is a nice shortcut to have available.

>I think the EU is going to be dependent on US tech (other than EUV lithography machines, very cool

Well, ASML's EUV light sources are based on licensed US IP from Sandia Labs, and manufactured in the US by CYMER, which ASML bought, but they still operate and manufacture out of California, so the EU is not sovereign/independent here (neither is any country).

This doesn't mean much anyway, since despite ASML being European, their machines all go to export and EU doesn't put any of those machine to good use domestically, with the most cutting edge semiconductor fabs on EU soil being the Germany based TSMC fabs on the much older 16 and 12nm nodes, far bigger than the 3nm that Taiwan and US operate domestically.

Re: Mistral built a $14B AI empire by not being American

#160

Post by vidarh »

dd8601fn wrote: post #67
vidarh wrote: post #48

Earlier quoted context omitted.

Unless/until there is a risk that the chips themselves are backdoored and trying to exfiltrate data, European companies that host in Europe still solve a big problem for use of certain data in Europe. It's not a purity test. Relying on US chips in not the same deal-breaker for all but the most extreme situation as relying on a poorly regulated US company to run the inference.

Out of curiosity, how does a chip that does inference calculations exfiltrate data without being seen? Has this happened already or is it just conceptually possible?

Not to my knowledge, and that's the point.

Though cards could if a provider has poor opsec. But I see no particular reason to worry about that either.

Post reply on HN