Mistral's been quiet lately...

[-]

Dark_Fire_12@reddit

Soon https://preview.redd.it/q54016w8a59e1.png?width=598&format=png&auto=webp&s=8b81656192203632a0b096b1538799be1d8f192a

Reply

[-]

SocialBudai@reddit

Mistral seems to be the one so far. They made me happy. It's like blizzard with Diablo 1.

Reply

[-]

silenceimpaired@reddit

I would almost be interested if Qwen didn’t have better performance and licensing across the board for my use.

Reply

[-]

Unless backed by hardware/infrastructure providers, it is very hard for a AI company to survive. Mistral have almost a billion in funding, but none of the investors are the likes of microsoft, google or amazon. Lately alibaba and google have significantly raised the bar. Facebook is also in the race, IBM has joined recently. All of them have open source or semi-open source models. So I think no one really cares about mistral at this moment, unless they pull off something comparable to o1+

Reply

[-]

zitr0y@reddit

> IBM has joined recently And their 2b model us surprisingly good. I was trying out a dozen models for a sentiment analysis task and theirs came a close second for that task after qwen2.5:3b (better than qwen2.5 7b, llama 3.1 8b and many more surprisingly)

Reply

[-]

Bitter-Good-2540@reddit

Which 2b model?

Reply

[-]

zitr0y@reddit

It is called granite3.1-dense

Reply

[-]

Bitter-Good-2540@reddit

Thanks! You tried to use it for local CPU rag?

Reply

[-]

zitr0y@reddit

No, I gave it a number (>200k) of German sentences with rapper names in them and made it categorize how positively or negatively the sentiment in the sentences is in regards to the rapper (only giving out a number between 1 and 5). I ran on GPU via ollama and its python integration. Feel free to ask more questions about it, I'm currently writing the research paper :D

Reply

[-]

Willing_Landscape_61@reddit

Did you compare with Bert models? Is seems to me that LLMs aren't the right tool for the job of text classification. (It's not like you are actually generating text).

Reply

[-]

zitr0y@reddit

You make a good point. In my class, it wasn't really made that clear what Bert actually does, I thought it was just an earlier, worse version of LLMs still used as a baseline in research. But it would likely have been a more fitting tool for the task. That said, qwen 2.5 3b did decently overall, with 65% perfect agreement and 95% off-by-one classification, zero shot.

Reply

[-]

FPham@reddit

Let's face it, once google realised they had the know how all the time, it went pretty well with Gemini...

Reply

[-]

Massive_Robot_Cactus@reddit

"Facebook is also in the race" Bruh.

Reply

[-]

nrkishere@reddit

yes they are. Nowhere said they are making state of the art models, but based on any leaderboard, LLAMA models are ahead of mistral's and this is the point of the argument.

Reply

[-]

Bitter-Good-2540@reddit

Llama is often used for fine tunes

Reply

[-]

FlerD-n-D@reddit

It's the other way around. He's saying you're understating what Facebook is doing.

Reply

[-]

Massive_Robot_Cactus@reddit

Yup, I could have been clearer. Just because Meta doesn't have a large cloud business doesn't mean they don't have one of the 5 largest data center footprints (and GPU compute) in the world.

Reply

[-]

thereisonlythedance@reddit

Mistral have provided the best all round local model in actual use (Mistral Large) and nobody cares about them? No. If nobody cared this thread wouldn’t exist.

Reply

[-]

silenceimpaired@reddit

Their licensing is a big speed bump for me and performance isn’t big enough to switch from Qwen and llama 3.3

Reply

[-]

fka_nate@reddit

But Mistral is literally funded by amazon? and Nvidia? might wanna do some more research on them lmao They’ve also released some really great models (esp for the gpu poor).

Reply

[-]

LevianMcBirdo@reddit

You know how much stuff these companies fund and how little goes to Mistral in the ai sector?

Reply

[-]

nrkishere@reddit

**Strategic partnership** is nothing like funding. Every cloud providers are partners of every open source models, because it brings them revenue. The primary investors, according to crunchbase are Andreessen-horowitz, salesforce and bnp paribas

Reply

[-]

Only-Letterhead-3411@reddit

I hope EU AI Act won't be the end of Mistral. I feel like Mistral really lost traction after that BS.

Reply

[-]

medialoungeguy@reddit

Their commitment to overregulating will be their last move.

Reply

[-]

lleti@reddit

It has been deeply deeply painful to watch us regulate ourselves into irrelevancy.

Reply

[-]

Feisty-Pay-5361@reddit

Well, on the flip side - EU (parts of it anyway) will also be the only place where realistically some form of UBI or monetary support for unemployment will happen. If mass job loss starts in the future due to all the unregulated AI rapidly advancing, citizens of US or Asia are absolutely screwed compared to Europeans (at least Nordic countries will for sure do ok, some others might join in).

Reply

[-]

CNWDI_Sigma_1@reddit

There is no money even for pensions, we are in a pension crisis. UBI would require at least 10x to 20x as much.

Reply

[-]

lleti@reddit

> the only place where realistically some form of UBI or monetary support for unemployment will happen. With what money tho We've sorta regulated outselves out of every major/cutting-edge industry, and a lot of our talent have left shore for the US or the Middle East to enjoy 4x the salary and 0.1x the taxes. Coupled with that, the Euro has been in steep decline against the USD since the financial crisis, with no sign of relief. Unfortunately I don't think there's gunna be anyone to actually pay for UBI on our shores.

Reply

[-]

Disastrous-Peak7040@reddit

JD & Elon are fans of UBI and minimum wage funded by innovation. They say "put kiosks in McDonalds, make more profit, pay better wages, build more restaurants". They're official on raising Fed min wage. The old school conservatives hate it. We may be entering a new pro-tech, pro-worker era?

Reply

[-]

lleti@reddit

Could be in the US The EU will likely regulate & fully outlaw anything which automates away a single job, even if it could fund thousands of UBI recipients in return. We've sorta wrecked ourselves in that regard tbh, we have bureaucrats who outlaw technology without even understanding the bare basics about it.

Reply

[-]

Feisty-Pay-5361@reddit

Well I am not saying the chances are good I am just saying proportionally it will likely be the only place where governments \*might\* give a fuck to come up with some system.

Reply

[-]

lleti@reddit

Yeah, I'm sure they'll try to do something or other - but it's just as likely to accelerate the collapse. Can't just print euros to feed the masses when there's nothing of value backing them.

Reply

[-]

procgen@reddit

Eh, a lot of Americans have 401ks and they'll be absolutely raking it in.

Reply

[-]

AssistBorn4589@reddit

> Well, on the flip side - EU (parts of it anyway) will also be the only place where realistically some form of UBI or monetary support for unemployment will happen. That is not flip side, that is getting fucked while getting fucked.

Reply

[-]

ohio_rizz_rani@reddit

Why thought? Isn't it better for them especially since the act itself talks about giving insights into data used and their models are open source. I think this is an advantage for then in the EU region it's also home grown company so I don't see why The EU AI act is a speed breaker.

Reply

[-]

Only-Letterhead-3411@reddit

The thing is EU AI Act is an hindrance put on Mistral's back in the AI race. While companies like OAI and Anthropic train their models on everything they can get their hands on, Mistral is forced to only use data they own themselves. These closed-source models are very good because they are trained on a lot of copyrighted data. I mean, previous year ChatGPT was giving people working windows license keys when asked. I think OpenAI is the proof that even the professional customers don't care about transparency and explainability, they care about quality and performance

Reply

[-]

goingsplit@reddit

It's plain obvious even in whisper to someone non-knowledgeable like me.. At the end of each inference the model spits a note about subtitles... Meaning it has been trained with copyrighted movies and subtitles produced by people

Reply

[-]

woutertjez@reddit

As someone that works in a large EU headquartered MNC, I can confirm accountability and transparency trumps model power. GDPR and other data/digital related acts in the EU are no joke when it comes to fines. We’re talking multiple percentage points of **global** turnover.

Reply

[-]

InertialLaunchSystem@reddit

What's wild to me is the EU thinking they could enforce this without risking a global trade war. Imagine trying to fine Apple or Nvidia at 4% of global *turnover* (not even profit) - that is a very fast way to end up on the losing side of a terrible trade war.

Reply

[-]

ohio_rizz_rani@reddit

What ever you say it's 100% valid, I agree we live in a capitalistic world. What I am also saying is that there are certain industries like finance, healthcare, pharma where transparency and explainability plays a huge role. Mistral still has a very good chance.

Reply

[-]

InertialLaunchSystem@reddit

If it prevents us from reaching AGI, copyright will have outlived its purpose and will cause more harm than good.

Reply

[-]

MorallyDeplorable@reddit

Copyright irself is unethical.

Reply

[-]

Many_SuchCases@reddit

That part in the EU AI act also means not breaking copyright, which is a question most companies aren't ready to answer. And the need to give insight is only a small part of the act.

Reply

[-]

GraceToSentience@reddit

Baseless Bs: "The AI Act **introduces limited exceptions for text and data mining**, recognizing the importance of balancing copyright protection with promoting innovation and research." [https://keanet.eu/eu-ai-act-shaping-copyright-compliance-in-the-age-of-ai-innovation/#:\~:text=The%20AI%20Act%20introduces%20limited,with%20promoting%20innovation%20and%20research](https://keanet.eu/eu-ai-act-shaping-copyright-compliance-in-the-age-of-ai-innovation/#:~:text=The%20AI%20Act%20introduces%20limited,with%20promoting%20innovation%20and%20research)

Reply

[-]

ohio_rizz_rani@reddit

I don't think it's necessarily bad , because companies like mistral will always have customers in heavily regulated industries where transparent and explainability plays a huge role.

Reply

[-]

SpargeOase@reddit

The customers are paying for the best models. You can't make the best models if you don't have the best quality data. 'Training data' transparency doesn't bring any benefits for most of the end users. We, Europeans, are just coping with this heavy regulation bullshit.

Reply

[-]

Nyghtbynger@reddit

What's sad is that Eurocrats **do believe** law makes money. In Luxembourg they do money with copyrights and when leveraging patent. That's an horrible way of doing money that will be made irrelevant in the few next years like the European Union It seems

Reply

[-]

brotie@reddit

https://i.redd.it/6hvsku7d579e1.gif

Reply

[-]

LevianMcBirdo@reddit

Exactly, having AI that complies with the rules gives you a giant market pretty much for yourself. Also the EU AI act is still not enforced yet (most stuff has a two year period, so 2026) and still Mistral is quiet now for months.

Reply

[-]

Any_Elderberry_3985@reddit

Maybe in Europe but the rest of the world including the US does not care about "the rules" as their is currently no legal risk outside of US IMO, Europe acted too quickly and likely gutted any development from Europe. Don't worry though non European companies will gladly gobble the data and train on it.

Reply

[-]

ThenExtension9196@reddit

Engineers came to America to get paid top dollar. Eu is no place to develop tech.

Reply

[-]

Nyghtbynger@reddit

Do you know how much an average data engineer/scientist is paid after taxes in France? 3500€ or 4000USD per month lol. Or make an effort. You can be top whatever, the state need to pay all the pensions from the old foggies and various welfare programs. That's the only country on earth where retirees earn more than working people. And no one is shocked when you tell them. Culture is good, food is good, cities are top class but doing business and working in France is one of the shittiest thing imaginable in the country, making the upsides *unaffordable*. People sometimes says that the US is a third-world country when it comes to catering to the people and the infrastructure. France is third-world when it comes to not being disappointed when starting an innovative project. No wonder the greatest minds are f*cking fleeing the country ~~I will too~~. Entrepreneur is a french word lmao

Reply

[-]

Original_Bend@reddit

3500€ a month after taxes is in the top tier for a data engineer, maybe in Paris.

Reply

[-]

4sater@reddit

Seriously? Wtf.

Reply

[-]

Nyghtbynger@reddit

And you need 4500€ after taxes to live comfortably in Paris...

Reply

[-]

4sater@reddit

Why the IT salaries are so low in France? I mean you can make 3500 euro after tax as a software/ML engineer even in some developing countries like China or Russia with much lower cost of living... is there a huge brain drain from France to the US and neighbouring countries like UK or Netherlands which afaik have higher salaries for skilled engineers?

Reply

[-]

Nyghtbynger@reddit

Being a software engineer in France is disregarded. Anyone touching a computer is basically an untouchable (indian Dalit). I've seen people that are geniuses in their field, not getting any job because they didn't do the right study or didn't stay long enough in their previous companies, or just don't know the right tool. People with lesser skills but some 6 months training in the tool via some training organization whose boss knows the hiring company have more chance of being hired. That's like a cartel where the manager are some nerdy assholes with a comp sci degree and they only hire asshole. If I seem salty it's because I had to face this kind of people that are average everywhere, except for their big egos. Now looking for options in "third world countries". In Thailand or Malaysia, I could earn 1:1 salary in euro compared to what you earn in smaller towns. That's approx 2.3 times better

Reply

[-]

4sater@reddit

Wow, that's shitty and, tbh, really stupid considering that huge chunks of economy are becoming more digital and are running on software engineers. Not to mention that France is just shooting itself in the foot in AI race... As a French citizen, could you try to go to Netherlands or perhaps Germany? They have better salaries for software developers & AFAIK you don't need a work visa since you are a EU member? Good luck!

Reply

[-]

Nyghtbynger@reddit

Thanks for your words. I have french and thainayionalities. 🤭 That's even better ! I tried in Germany (I speak German) but their GDP dropped 10% since they don't have cheap energy anymore. They don't really hire foreigners right now... I still believe the current situation in France is a waste. The country basically have free nuclear energy and produces a lot of scientists (in data too). Having good AI should be an evidence. But I must be realistic, France is in cultural decline for one century now. They can't imagine themselves without being a global power that relies on a long gone colonial hinterland. Time for the big questioning and some practices changes.. If that doesn't work I can still changes carrier and become a plumber. Fine by me 🤷‍♀️

Reply

[-]

Josh_j555@reddit

> Culture is good, food is good, cities are top class This is quickly changing as well, sadly not for the better.

Reply

[-]

MorallyDeplorable@reddit

> 3500€ or 4000USD per month Is that supposed to be impressive?

Reply

[-]

Bitter-Good-2540@reddit

It's a place to develop. Get you degree for cheap. Got government funding and move to USA. Blackforest did that lol

Reply

[-]

GraceToSentience@reddit

The EU AI Act is a self reported thing much like AI regulations in the USA People don't know what it does and think it's some kind of tough regulation. It's not.

Reply

[-]

mahithefish@reddit

Yeah as much as I’m behind that Act, it’s a very very tough constraint for them to remain competitive but maybe they can win in EU.

Reply

[-]

anonynousasdfg@reddit

In the worst case with enough investor support they may move their headquarters to U.S, although I'm not sure if it will help them in the long run to become an independent company without being acquired by some closed-source property giants or will just make them bankrupt.

Reply

[-]

lilythompsilly@reddit

Pixtral 12B released recently is quite awesome too. It is exceptional at doing OCR and interpretation related tasks.

Reply

[-]

Willing_Landscape_61@reddit

They probably are frantically reading the DeepSeek 3 paper right now!

Reply

[-]

BetEvening@reddit

daily reminder: https://preview.redd.it/1wwc7gqpkc9e1.png?width=968&format=png&auto=webp&s=c5c80740aac85c931c2afd856dc62400a78b422a :

Reply

[-]

PrinceOfLeon@reddit

Mistral is currently in the process of opening a Bay Area office. I wonder if they'll incorporate separately there in order to get around the EU's restrictions on AI. Personally I lost interest in following them after they stopped releasing under open licenses.

Reply

[-]

Nyghtbynger@reddit

I would have opened a Shenzen or Texas office instead

Reply

[-]

MoffKalast@reddit

EU: It's treason then.

Reply

[-]

Ok_Wear7716@reddit

It’s time for annual 8 week holiday in France, so it makes sense

Reply

[-]

Zangwuz@reddit

Not really, Pixtral Large was released just one month ago.

Reply

[-]

ninjasaid13@reddit

yeah but they've never been quiet for a whole month besides this january and last year october and november.

Reply

[-]

FortranUA@reddit

And need to say it's quite good, in some moments even better then gpt4o in describing images

Reply

[-]

infernys20@reddit

*than

Reply

[-]

FortranUA@reddit

Thanx 😊 I was very sleepy when I wrote

Reply

[-]

JzTheLazy@reddit

Gn mate 😴😴

Reply

[-]

umarmnaq@reddit (OP)

That's a millennium in AI time

Reply

[-]

TheDreamWoken@reddit

Ok

Reply

[-]

MoffKalast@reddit

It has been 84 years...

Reply

[-]

Such_Advantage_6949@reddit

I dont even know what you complain about. Why not asking meta and google to release more who also have more resources? Mistral released pixtral large just recently. Whereas meta and google both doesnt release too end model. The only company that released more is alibaba with their qwen series.

Reply

[-]

Healthy-Nebula-3603@reddit

Google yes . But meta announced llama 4 and soon will release also 2 weeks ago released the llama 3.3 model.

Reply

[-]

Such_Advantage_6949@reddit

Dont think llama 4 released date is confirmed yet right. For 3.3 is more incremental update, whereas their 3.2 vision part is not as good as competitor. In comparison, qwen released good vision model and reasoning model. Pixtral have good vision capabilities. To be honest, i am sure they are capable of release something better. But it feels like the bigger player is intentionally holding back

Reply

[-]

kif88@reddit

Last I heard was Mark Zuckerbergs video on Facebook. He said llama 4 should be out "in 2025"

Reply

[-]

Such_Advantage_6949@reddit

Out in 2025 is not exactly a release date lol

Reply

[-]

Healthy-Nebula-3603@reddit

If you read recent papers from meta and if they implemented that in llama 4 ... then will be wild 😅

Reply

[-]

Dark_Fire_12@reddit

It's a bit, they do ask for them as well. The rotation is Mistral > Meta (Llama) > Google (Gemma) > Cohere. We got 3.3 from Meta and a new updated Paligemma from Google, as well as a 7B from Cohere. Mistral is next up.

Reply

[-]

Such_Advantage_6949@reddit

Haha i nvr realise there is a circle of release lol. Lets see

Reply

[-]

Mother-Ad-2559@reddit

Let them cook

Reply

[-]

FantasticRewards@reddit

Mistral Large is still my favorite model but would love a new Miqu (70B).

Reply

[-]

martinerous@reddit

A Mistral-not-so-small-and-not-that-large would be nice. 32B is the sweet spot for me. I really like the current Mistral Small model for its overall consistency when prompted to follow long step-by-step interactive scenarios. In comparison, other models (even Qwen 32B) mix up the steps or items or interpret the instructions in abstract manner. Mistral Small is the most solid, but +10B would benefit it, I think.

Reply

[-]

johnny1064@reddit

Maybe Mistral’s busy cooking. Unlike OpenAi who’s too busy talking, then gets their ass handed to them by Google. (Veo2)

Reply

[-]

Spammesir@reddit

I get your point about SORA but o3's definitely good

Reply

[-]

EquivalentIce7442@reddit

How do you know? They didn't release it..

Reply

[-]

procgen@reddit

How do we know? The benchmarks results, obviously.

Reply

[-]

Few_Painter_5588@reddit

Those benchmarks were flubbed by basically giving the model infinite time and resources to think.

Reply

[-]

procgen@reddit

That's either a misunderstanding on your part or a blatant lie: https://arcprize.org/blog/oai-o3-pub-breakthrough Time per task was ~13 mins on the semi-private eval.

Reply

[-]

EquivalentIce7442@reddit

Ahahah yeah we'll see when and if..

Reply

[-]

procgen@reddit

What do you mean? Francois Chollet already confirmed it, lol.

Reply

[-]

EquivalentIce7442@reddit

I'd like to try it then we'll see..

Reply

[-]

procgen@reddit

The fact remains that no other model has come close on the ARC-AGI or frontier math benchmarks. The reason you can't use it now is because it's absurdly expensive to run, but the costs will drop fast.

Reply

[-]

squareOfTwo@reddit

did you try it? The answer is no

Reply

[-]

procgen@reddit

o3 is outperforming humans on ARC-AGI, lol. They have the most powerful research model that's been publicly revealed.

Reply

[-]

fallingdowndizzyvr@reddit

I think they have to tread lightly and carefully with the new EU regs to worry about.

Reply

[-]

candre23@reddit

Not really. They dropped a new version of the 22b in September. October was a new 8b. A month ago we got two new versions of largestral - with and without image support. I know this space moves fast, but going one whole month without a new model is hardly "sleeping".

Reply

[-]

kif88@reddit

And that 3b ministral they keep behind API.

Reply

[-]

ForsookComparison@reddit

dying for a new codestral

Reply

[-]

pigeon57434@reddit

you know who has really been totally silent? Anthropic. I wonder what they will do Claude 3.5 was a fucking beast but they havent released the next gen models yet and are behind now

Reply

[-]

Any_Elderberry_3985@reddit

They released pix large not long ago. They don't get much press anymore because there are other good models and they have no commercial use without licensing.

Reply

[-]

silenceimpaired@reddit

Came here to say this. My interest in them died the moment they switched to a license like this… especially since their dataset is probably based off the work of others without their consent.

Reply

[-]