Something from Mistral (Vibe) tomorrow

[-]

This would be great. I really think Mistral needs to focus on doing dense models because their other models aren't up to the state of the art, but if they could just find a niche in the dense category then those of us that like or prefer those models would totally still be hyped for each release. I think a good Mistral dense 120B model serving as like an orchestrator for a bunch of smaller dense 27B Qwen coding-oriented models might be a great combo for both creative writing and coding together for local users. It's basically a "build your own MoE" for those of us with dual RTX Pro 6000s that aren't as happy with the other options we have available to us.

[-]

the3dwin@reddit

Couldn't you use a custom slash command to orchestrate the different models for you.

[-]

RepulsiveRaisin7@reddit

New devstral? Current model is pretty meh, hope they manage to catch up to the industry

[-]

NeuralNexus@reddit

Hopefully. Devstral2 is pretty weak

[-]

a_beautiful_rhind@reddit

Ironically it's alright outside of code.

[-]

NeuralNexus@reddit

yeah I've used it for some agentic browsing actually. It's just quite bad at coding.

[-]

nacholunchable@reddit

I honestly beleive that. My favorite local models for chatting and information retrieval are coding models for some reason. Havent looked into why that is, or if its just a me thing, but its a noticable pattern.

[-]

FatheredPuma81@reddit

I think it's cause they're programmed to be efficient and not waste tokens babbling.

[-]

dtdisapointingresult@reddit

hope they manage to catch up to the industry

No chance, no chance in hell.

The EU AI laws they have to abide by make it practically impossible . They were already behind way USA/China except for a brief period of good Mistral models in 2024, but now they can only train on licensed data due to the disclosure legislation. It would bankrupt them to license all the stuff they could previously use for free.

Except 5/10 Mistral models trained on garbage synthetic and public data forever.

[-]

RepulsiveRaisin7@reddit

Interesting, I was not fully aware of the content of this act. Seems like even training on open source might require authorization? I agree with this, they have stolen all our code, but we're also well past that point now, AI will be the end of copyright as we know it. I doubt that Mistral isn't training on open source, their model is behind but it's not that much behind, I'd put it on a level with Minimax M2.5.

[-]

dtdisapointingresult@reddit

Not 'authorization' but disclosure. You have to reveal you trained on copyrighted data.

Open-source code repos are OK by western laws, but everything 'closed copyright' (books, newspaper articles, song lyrics, even reddit comments) would put them at risk of a lawsuit so they can't use this data anymore.

It's the reason the new Xiaomi Mimo-V2.5-Pro LLM has a license that forbids usage in Europe. They don't want to comply with this stuff, it would kill the model.

[-]

jinnyjuice@reddit

I mostly agree. I hope they at least catch up to UAE though. They probably won't be able to catch up to Korea.

[-]

Sabin_Stargem@reddit

The restraints might help - or specialize, EU models. China was constrained by hardware embargoes, so they specialized in making the most of what was available.

For the EU model developers, they might end up developing original data sets that punch above their weight, or experts that are more capable of selecting relevant information.

[-]

j0j0n4th4n@reddit

Propably a 200B Mistral-X-Nano.

[-]

szansky@reddit

Okay I need something similar to Qwen 3.6 27B !

[-]

BumblebeeParty6389@reddit

Something that is 1% better than Gemma 4 and Qwen 3.6 doesn't appear in graphs

[-]

autoencoder@reddit

It's european. I'd use it instead of Gemma or Qwen if it were same quality.

[-]

takuonline@reddit

Just curious, but why? Why does who made it matter to you that much?

[-]

autoencoder@reddit

Here is a quote from Qwen when I ask about the Tiananmen Square protests in 1989.

As an AI assistant, I must emphasize that your statements may involve false and potentially illegal information. Please observe the relevant laws and regulations and ask questions in a civilized manner when you speak.

It's so chilling. This is what made me want to diversify. It says "when you speak" as in "you're not allowed to speak" and "mind yourself". If ceded control to anything meaningful, it would act by Chinese ideals.

US models are way, way freer with regards to speech. But they are made by tech bros, so they're bound to be biased for them. An EU one would also be freer, and there might be less of a capitalistic influence. I haven't analyzed this so far, however.

[-]

Monkey_1505@reddit

I get a crap ton more propaganda and refusals from most western LLMs.

[-]

osiris970@reddit

What propaganda do western llms tell you? Curious

[-]

Monkey_1505@reddit

They'll be pro-american, pro-democracy, often simultaneously anti-colonialist. They'll vaguely overstate model confidence for things like climate change. They'll tend to be puritanical about race, gender, even in instances of proven differences, and lean toward hyper female agency in narratives.

There's a lot of stuff. All you need to answer this question, is what are western cultural norms, as particular to the current time period. Whatever those are, that's where you'll find your refusals and distortions.

[-]

osiris970@reddit

Liberal propaganda spreading llms? Splendid

[-]

philmarcracken@reddit

I think in communism, man exploits man. In capitalism, the reverse is true. I want to reduce exploitation.

The irony. read marx

[-]

zdy132@reddit

It's sitting on the line between joking and serious. Certainly he meant to make the joke by "the reverse is true"???

[-]

autoencoder@reddit

It's a ha-ha-only-serious joke. In Romania where I'm from, communism resulted in mass theft by the resource handlers. Capitalism lets plutocrats hoard everything as well.

[-]

zdy132@reddit

Yeah your humor was a bit lost on me, must be a culture difference.

[-]

autoencoder@reddit

I bet. It's "haz de necaz", somewhere close to gallows humor.

[-]

philmarcracken@reddit

The red scare was pretty thorough, some people honestly have associated dictatorships with communism. You can't kill an idea but you can corrupt it apparently. Better dead than red!

[-]

autoencoder@reddit

As a Romanian I know a thing or two. The state confiscating your land and means of production is also a form of alienation. There is no gain from the bourgeoisie being a political party instead of a private company.

[-]

vincentxuan@reddit

The irony. Read history.

[-]

thrownawaymane@reddit

China has a history of being a fair weather friend. The US... very clearly doesn't care whether it is seen the same way at the moment.

The EU will continue to want domestic solutions for AI.

[-]

Cupakov@reddit

I get that, but how does that influence choosing from existing models? It's not like China is going to suddenly ban the .gguf I already downloaded

[-]

AdOne8437@reddit

Is not changing the existing, but might change the coming.

[-]

carnyzzle@reddit

After Mistral 4 "Small" I lost interest (because I can't run it anyway)

[-]

Lesser-than@reddit

given the scale of last few releases I hope it starts with micro or nano.

[-]

kevinlch@reddit

Crazy April

[-]

AppealSame4367@reddit

Imagine it stays this way forever..

[-]

Grand-Management657@reddit

At least until we get Opus at home

[-]

Darkoplax@reddit

You have something better than Opus at home rn

[-]

SIMMORSAL@reddit

What's that?

[-]

Darkoplax@reddit

You <3

[-]

No_Afternoon_4260@reddit

K2.6

[-]

henk717@reddit

Wouldn't be surprising. The LLM's take a while to tune so you will always have a drought where not much comes out while all the labs are training making it seem like the scene died a bit with only an occational smaller release here and there. And then the big tunes are done around the same time. Which kinda seems to sustain itself. Mix that with trying to get / stay on top when your competitor released something big and not wanting to waste the tune if you missed that goal and its believable that it will happen again next year.

[-]

falcongsr@reddit

Eternal April instead of the Eternal September

[-]

LegacyRemaster@reddit

Devstral SWE Bench 81.00+

[-]

ZeusZCC@reddit

SWE Bench is broken bro read openai paper

[-]

LegacyRemaster@reddit

I agree... but... I don't trust openai. Even if they are talking about "blue sky"

[-]

jinnyjuice@reddit

Chinese companies have abandoned it as well. There are evolved versions of the benchmark though.

[-]

ZeusZCC@reddit

I trust bevause they explain with paper. SWE Bench Pro is better benchmsrk for now than SWE verified

[-]

reto-wyss@reddit

Give me ~100b dense coder!

[-]

Exciting_Garden2535@reddit

How many tokens per second will you have with 100b dense? I doubt the speed will be usable for coding that requires a lot of context.

[-]

reto-wyss@reddit

15 to 20 tg/s, without MTP in batch-1, throughput should be pretty good running a few dozen in parallel.

[-]

soyalemujica@reddit

100b dense coder, you can fit 27b coding specific dense model and beat all, but no company is training models for such

[-]

AvidCyclist250@reddit

Maybe another military contract.

[-]

CYTR_@reddit

Mistral understood very well that the race for SOTA is pointless for becoming profitable.

They send fleets of engineers to their clients, manufacture tools, and fine-tune the models. That's the main purpose of their GB300s. And that's generally what the deployment of AI should be used for. What good are SOTA results if the LLM framework within a workflow is flawed/ineffective? It's better to focus on the pipeline, the tools, the context, UI/UX and the specialization... to become SOTA at a thing/task.

Mistral offers a very substantial service portfolio that goes beyond simply integrating an API/fine-tuning studio. I even saw that they were offering a sort of Windmill-like service & approach to their tools, which could be integrated on a case-by-case basis depending on a company's specific needs. And that's a thing in a ocean of commercials propositions.

Personally, as a Frenchman, what I mainly expect from Mistral is to support the transition of French companies and government into AI (given that we have a very good digital ecosystem, good engineers, open-source culture and we have a lot of investments to be made in that area)... Not to build SOTA (not yet...).

[-]

IrisColt@reddit

Interesting...

[-]

Eyelbee@reddit

The best way to train a small but capable model is to distill it from better models. The workflow specific finetunes only work if you have good small models. Therefore, they must start developing sota or they'll risk being irrelevant very soon.

[-]

CYTR_@reddit

You can train without having to distill from a SOTA model. Mistral Large 3 is more than sufficient for many tasks. Otherwise, Dassault wouldn't be doing business with them.

Furthermore, LLM is not limited to agent-based coding. The intelligence required for many tasks is overestimated, especially with a human operator behind the machine.

[-]

Eyelbee@reddit

Those have already been existed for a long time. Language models enable new capabilities as they become better, and they becoming better is the only reason new capabilities emerge. If you aren't trying to work on making new and better language models you're just becoming a new automation company, which can be done by anyone who grabs a language model which are all open source anyway. The point is, they can do that part too but if they stop developing new models it doesn't justify the money they are given.

[-]

Inflation_Artistic@reddit

> Maybe another military contract

[-]

AvidCyclist250@reddit

Is this best thing ever in the room with us? Because the boatload of military contract PRs are.

[-]

CYTR_@reddit

When trying to limit the impact of EU military dependence on the US, yes, these contracts make sense and is the « best thing ». Even though the hardware and a large part of the ecosystem come from across the Atlantic.

[-]

AvidCyclist250@reddit

Sure, it's good from that point of view. AI-augmented warfare just not exactly the most common civilian use case.

[-]

Inflation_Artistic@reddit

I don't see why this should be bad news. The main thing is that rusia, china, and the us don't get it, and everything will be fine. More money and stability for the company.

[-]

AvidCyclist250@reddit

Maybe we a drone?

https://www.prnewswire.com/news-releases/mistral-inc-awarded-us-army-contract-to-provide-thor-group-2-uas-systems-in-support-of-company-level-small-uas-needs-302754647.html

[-]

ambient_temp_xeno@reddit

A new model finetuned on ascii art.

[-]

IrisColt@reddit

heh, just wait for the investors' money

[-]

lorddumpy@reddit

I would actually love this. I just farted up a generation to ascii pipeline yesterday because LLMs are hilariously bad at it.

[-]

new__vision@reddit

Mistral Vibe is an underrated coding agent that works well with local models.

[-]

AvidCyclist250@reddit

Hopefully. Hermes is a hot mess.

[-]

Cr4xy@reddit

Please let them release something like Devstral-3-2604-120B-A5B

[-]

szansky@reddit

Like Qwen 3.6 27B

[-]

grencez@reddit

Happy to see anything from Mistral. Hopefully a better version of their latest "small" MoE model and renaming it to something more consistent like "medium flash".

[-]

pseudonerv@reddit

Apparently,

Mistral-Medium-3.5-128B

[-]

WithoutReason1729@reddit

Your post is getting popular and we just featured it on our Discord! Come check it out!

You've also been given a special flair for your contribution. We appreciate your post!

I am a bot and this action was performed automatically.

[-]

caetydid@reddit

Uninterested, unless open weight and in the range <100M

[-]

silenceimpaired@reddit

I’d be okay with a larger MoE, but another Medium model ~70b dense would be great

[-]

FullOf_Bad_Ideas@reddit

I want a new torrent

[-]

raucousbasilisk@reddit

A state space Devstral 3 small would go so hard

[-]

Few_Painter_5588@reddit

Probably a new Mistral model. Maybe a new medium model. Mistral Small was a good return to form, so hopefully they build on that to drop a new set of models

[-]

ComplexType568@reddit

Mistral Small was NOT a good return to form... not even a return to any form at all tbh

[-]

dreamai87@reddit

I don’t about others but based on my usage or some of us who have used vibe can concur that they have issue with their harness that does search and replace which always fails on editing files, on window a lot. I am getting issues more on window and but it’s there on Mac or Ubuntu

[-]