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

Notes from the Mistral AI Now Summit

koenvangilst.nl

Post reply on HN
Print view View original

21–30 of 230 posts

Re: Notes from the Mistral AI Now Summit

#21

Post by lettergram »

trouve_search wrote: post #9

OK, I'm 100% rooting for both Mistral and task focused small models. But Mistral has fall really far behind since 2025Q3. It seems they can't get good reasoning models working at even medium context sizes, which is necessary to be at the table right now. Gemma4 and Qwen3.6 are currently best in the small size; Mistral's "small" model has ~4x the parameter count at 120B and isn't even competing with models a quarter i…

We actually found the Mistral Small 4, quantized to 4bit was comparable to Qwen 3.6 27B and is roughly the same size. At least from our experience on our use cases, the quantization of the Mistral model worked far better than trying to quantize the Qwen family.

Fully agree to your point though, Mistral in general is far behind where I'd expect and Qwen in particular is crushing it at the smaller sizes.

Personally, I'd consider anything 20B params and above a "medium" model. Small being 100B. I think obviously we can get to the huge 1-2T param models, but frankly the margin of accuracy improvement for the speed hit is kinda insane (1-2% for many metrics).

Re: Notes from the Mistral AI Now Summit

#22

Post by ogou »

I've said it before that Mistral is underrated. They are looking at real world use of LLMs and tooling. Bespoke models are very appealing to lots of non-tech centered companies and state agencies. Also, Mistral's actual platform is useful. While others are watching performance leaderboards like this is some eSports stream, they are building real world uses.

Re: Notes from the Mistral AI Now Summit

#23

Post by kriro »

doctorpangloss wrote: post #15
simonw wrote: post #3

> BNP Paribas runs Mistral models on-prem for KYC in Belgium, with sensitive data staying within the bank's walls. Abanca is using agent orchestration to handle sensitive customer information at a huge scale (2 million customers in their app). For European companies in regulated industries, this is a good alternative to relying on US hyperscalers. Mistral leaning into on-prem and European-hosted models is very smart.

Yeah but why use mistral on premises instead of Qwen?

We're talking about enterprise customers. The trivial answer is Mistral has sales teams and consultants from the same company that builds the models and from the EU.

Re: Notes from the Mistral AI Now Summit

#24

Post by throw14082020 »

simonw wrote: post #3

> BNP Paribas runs Mistral models on-prem for KYC in Belgium, with sensitive data staying within the bank's walls. Abanca is using agent orchestration to handle sensitive customer information at a huge scale (2 million customers in their app). For European companies in regulated industries, this is a good alternative to relying on US hyperscalers. Mistral leaning into on-prem and European-hosted models is very smart.

Respectfully, I don't think it's "very" smart. It is a fair option given their limited options? Everyone is doing FDE or (customer engineering to be more transparent) because otherwise they will just be seen as markup on token cost. And the Neo-SaaS companies will take the money instead.

Who else will buy their AI?

and what other options do they have?

Re: Notes from the Mistral AI Now Summit

#25

Post by joe_mamba »

simonw wrote: post #17
doctorpangloss wrote: post #15

Earlier quoted context omitted.

Yeah but why use mistral on premises instead of Qwen?

One reason might be that Mistral doesn't have a risk of weird training biases that were required by the Chinese government.

>weird training biases that were required by the Chinese government

What is "weird training biases" to us might not be weird to them and vice versa. Just ask the Chinese what they think about LGBTQ+, Japanese, pride parades, Islam and colored minorities.

Every nation has its own biases injected in its domestic LLMs at this point. Otherwise they risk getting in trouble for hate speech/disinformation in the jurisdiction where they operate.

Same how Google Maps cleverly biases the lines of disputed borders based on where you are viewing it from. Or how Google maps switched 'Gulf of Mexico' to 'Gulf of America' in an instant when the orange man signed the paper. Google won't want to anger the US administration the same way how Mistral won't want to anger France and the EU, so Mistral will have all the EU prime directives injected into its LLMs no matter if they're ludicrous or not. The law is the law whether you agree with it or not. Companies want to survive and will pander to whatever the whims the regime they live under are at the current moment regardless of what is right or wrong.

But if I'm using a LLM for personal projects or generating a photorealistic choreographed fight between Tom Cruise and Brad Pitt, I don't care what its political biases are, I care if it solves my problem better and cheaper than the competition, and here the Chinese models could end up winning the consumer market, which is why you see Mistral and other EU alternatives focusing exclusive on B-2-B corporate market.

Re: Notes from the Mistral AI Now Summit

#26

Post by ahnick »

echelon wrote: post #16
trouve_search wrote: post #9

OK, I'm 100% rooting for both Mistral and task focused small models. But Mistral has fall really far behind since 2025Q3. It seems they can't get good reasoning models working at even medium context sizes, which is necessary to be at the table right now. Gemma4 and Qwen3.6 are currently best in the small size; Mistral's "small" model has ~4x the parameter count at 120B and isn't even competing with models a quarter i…

Nobody trying to compete with Google, OpenAI, and Anthropic should be playing the small models / local models game. Foundation model labs should be building very large reasoning models, then leaving it to the community to distill them down. You can't scale a small model up, but you can scale a small model down. I'm convinced the only way we'll have a seat at the table in the future and avoid total runaway takeoff is…

I thought distillation meant small models don't have to compete with the big models and can always eventually achieve close parity, but it's just a matter of time to do the distillation? (i.e. how much lag do you want to live with) Am I oversimplifying?

Re: Notes from the Mistral AI Now Summit

#27

Post by antirez »

I really want Europe to be part of the AI development and research. And I strongly cheered for Mistral. But they are accumulating too much technological delay. This needs to be fixed, otherwise it will turn into yet another proof we are not able to run large tech with good results. Basically any Chinese lab is doing much better. It's not Mistral that created I don't want to say DeepSeek, but MiMo 2.5, Minimax 2.7, and so forth. There are only weaker and/or larger and slower (no MoE) models. Not good.

Re: Notes from the Mistral AI Now Summit

#28

Post by baq »

trouve_search wrote: post #9

OK, I'm 100% rooting for both Mistral and task focused small models. But Mistral has fall really far behind since 2025Q3. It seems they can't get good reasoning models working at even medium context sizes, which is necessary to be at the table right now. Gemma4 and Qwen3.6 are currently best in the small size; Mistral's "small" model has ~4x the parameter count at 120B and isn't even competing with models a quarter i…

agreed, the next price increase from frontier labs (and the inevitable limits decrease in subscription tiers) will have people thinking real hard about their model providers and that's when mistral should be ready. however, given their recent performance, I realistically don't have my hopes high up.

Re: Notes from the Mistral AI Now Summit

#29

Post by farley13 »

echelon wrote: post #16
trouve_search wrote: post #9

OK, I'm 100% rooting for both Mistral and task focused small models. But Mistral has fall really far behind since 2025Q3. It seems they can't get good reasoning models working at even medium context sizes, which is necessary to be at the table right now. Gemma4 and Qwen3.6 are currently best in the small size; Mistral's "small" model has ~4x the parameter count at 120B and isn't even competing with models a quarter i…

Nobody trying to compete with Google, OpenAI, and Anthropic should be playing the small models / local models game. Foundation model labs should be building very large reasoning models, then leaving it to the community to distill them down. You can't scale a small model up, but you can scale a small model down. I'm convinced the only way we'll have a seat at the table in the future and avoid total runaway takeoff is…

I do think there's a chance open weight models have a bit of a moment with the costs of frontier models growing on business balance sheets. It's unfortunate from my "privacy loving" PoV that it's mostly Chinese models filling the gap. ( the top models on openrouter for instance ).

I have used Mistral models out of pure ideology for web agents and the like which aren't doing a lot of heavy lifting.

Re: Notes from the Mistral AI Now Summit

#30

Post by simonw »

joe_mamba wrote: post #25
simonw wrote: post #17

Earlier quoted context omitted.

One reason might be that Mistral doesn't have a risk of weird training biases that were required by the Chinese government.

>weird training biases that were required by the Chinese government What is "weird training biases" to us might not be weird to them and vice versa. Just ask the Chinese what they think about LGBTQ+, Japanese, pride parades, Islam and colored minorities. Every nation has its own biases injected in its domestic LLMs at this point. Otherwise they risk getting in trouble for hate speech/disinformation in the jurisdictio…

> What is "weird training biases" to us might not be weird to them and vice versa.

I agree. That's why I think European companies might prefer a European model.

Post reply on HN