Hacker News

I got really lost on their site, but to help a bit according to their model page

devstral-2512 devstral-latest and devstral-medium-latest are all devstral 2 https://docs.mistral.ai/models/devstral-2-25-12

labs-devstral-small-2512 and devstral-small-latest are devstral small 2

devstral-medium-2507 is devstral 1.0

and devstral-small-2507 is devstral small 1.1

I had the same experience. It's even more confusing when you want to create an API key because they are separated by product, maybe?

I have a general impression they are not interested too much in individual devs and making it suite their workflow. They want to be a B2B company and deliver a custom workflow per company.

Or it can just be a Google like problem where a big company one part doesn't talk to the other.

>data staying in the EU

This is really why Mistral has any support.

The models are bottom barrel, but its the best Europe has...

Although you could use Chinese models on European servers.

I also think that this is the best approach for businesses wanting to adopt AI to automate, streamline, etc their business.

The problem they have is that this is not a moat - their approach is easily reproducible.

If they can pull ahead in having the most number of pre-trained models (one for this ERP, one for that CRM, etc) and then being able to close sales to companies using these products and sell them on post-trained (give us your specific ERP customisations and we'll give you access to a model that is tailored to your business), then THAT is a moat.

But they need to do this without fanfare. Just close sales, and keep closing, basically. After all, even if other AI providers copy the process, the moat would already have been established for Mistral.

> Their emphasis on bespoke modelling over generalized megaliths will pay off.

Isn't the entire deal with LLMs that they are trained as megaliths? How can bespoke modelling overcome the treasure trove of knowledge that megaliths can generically bring in, even in bespoke scenarios?

The companies I work want onprem models, and no Chinese ones. Does mistral support onprem? ( For a price)

Agreed. I’ve used their platform to train smaller, specialized models. Something I could have done in Codelab or some other tool, but their platform allows me to just upload a training set and as soon as it finishes I have a hosted model available at an endpoint. It obviously has some constraints compared to running the training yourself, but it also opens up the opportunity to way more people.

Indeed, but even for coding use cases, Vibe is more of a focused “refactor/ write this function” aid than “write me an app” and it can work locally. For me that’s a lot more valuable as an accelerator to my workflow where the developer stays in control and fully involved in the process.

I agree. Just started using it. Can you give some examples of fields you maybe even prefer Mistral?

I use a pretty lightweight local Mistral model in LM studio for both creative and technical writing/iterating and it’s fantastic.

Yes, since it's not American, it will be the de-facto choice for most big European companies.

Is this the best Grok alternative?

If you couldn't use the words Europe to describe why you'd chose Mistral, you'd have no good reasons to choose Mistral.

Its just not good. Its bottom floor for LLMs.

I found it to be the best model if you want to talk about topics philosophical. It has no problems going deep and technical while other models tend to be afraid of overshooting the comprehension of the reader.

their ocr model is goated

also offering support for local deployments

Go Mistral !

first, there was .ai

next, it sounds like it's going to be .eu

but what about ai.eu

Pre-training mean exposing an already-trained model to more raw text like PDF extracts etc (aka continued pre-training). You wouldn't be starting from scratch, but it's still pre-training because the objective is just next token prediction of the text you expose it to.

Post-training means everything else: SFT, DPO, RL, etc. Anything that involves things like prompt/response pairs, reward models, or benefits from human feedback of any kind.

Probably marketing speak for full fine-tuning vs PEFT/LoRA.

I would guess:

Pre-training: refining the weights in an existing model using more training data.

Post-training: Adding some training data to the prompt (RAG, basically).

I think they are referring to “continued pretraining”.

I can imagine that, as usual, you start with a few examples and then instruct an LLM to synthesize more examples out of that, and train using that. Sounds horrible, but actually works fairly well in practice.

Probably just means SFT fine-tuning a base model, vs behavioural dpo and/or SFT fine-tuning a instruction model.

I'm thinking stuff like this:

https://denverite.com/2026/03/12/ai-recycling-facility-comme...

You could take a model like the one referenced in the article, retool it with Forge for oh I don't know, compost, and use it to flag batches that contain too much paper for instance.

These kinds of applications would work across industries, basically anywhere where you have a documented process and can stand to have automated oversight.

You can fine tune small, very fast and cheap to run specialized models ie. to react to logs, tool use and domain knowledge, possibly removing network llm comms altogether etc.

rag basically gives the llm a bunch of documents to search thru for the answer. What it doesn't do is make the algorithm any better. pre-training and fine-tunning improve the llm abaility to reason about your task.

RAG is dead

> After all, real humans – even geniuses – are trained on much much less data than the whole Internet.

It's certainly different data, but one could argue that real humans have been trained on 3.5 billion years of evolution data.

I was under this impression as well - I'd love to hear from someone who's deeper in the know about this!

Typical stages of training for these models are:

Foundational:

- Pretraining - Mid/post-training (SFT) - RLHF or alignment post-training (RL)

And sometimes...

- Some more customer-specific fine-tuning.

Note that any supervised fine-tuning following the Pretraining stage is just swapping the dataset and maybe tweaking some of the optimiser settings. Presumably they're talking about this kind of pre-RL fine-tuning instead of post-RL fine-tuning, and not about swapping out the Pretraining stage entirely.

"External Storage" whatever that is can not be the same as continous learning as it does not have the strong connections/capture the interdepencies of knowledge.

That said I think we will see more efforts also on the business side to have models that can help you build a knowledge base in some kind of standardized way that the model is trained to read. Or synthesize some sort on instructions how to navigate your knowledge base.

Currently e.g. Copilot tries to navigate a hot mess of a MS knowledge graph that is very different for each company. And due to its amnesia it has to repeat the discovery in every session. No wonder that does not work. We have to either standardize or store somewhere (model, instructions) how to find information efficiently.

What do you mean when you say "external storage?"

Ironically that was also the past of AI. In 2016 it was all about specialized models (not just training data, everything including architecture and model class/type) for specific tasks and that's the way things had been for a long time.

Are you suggesting that it's an aberration that from ~2019 to ~2026 the AI field has been working on general intelligence (I assume this is what you mean by "achieving benevolent knowledge")?

Personally I think it's remarkable how much a simple transformer model can do when scaled up in size. LLMs are an incredible feat of generalization. I don't see why the trajectory should change back towards specialization now.

I don't think that's true. Nothing points to specialized LLMs being better. General purpose LLMs are just much more useful in daily work.

My sense is that it sounds amazing in theory to executives who have never had to themselves look at internal data. In reality the internal knowledge base is a mix of incomplete, inaccurate, self serving lies, out of date and so on. At worst, the data is explicitly biased to hide reality from executives so the AI will look extra good to executives. Of course, a business that makes all tactical decisions based on lies is not going to do well.

Interesting to see. I thought they were promoting fine tuning

It seems to me that it is broadly the same thing, except they give you the resources to do it and expert knowledge.

lol

> Mistral AI has already partnered with world-leading organizations, like ASML, DSO National Laboratories Singapore, Ericsson, European Space Agency, Home Team Science and Technology Agency (HTX) Singapore, and Reply to train models on the proprietary data that powers their most complex systems and future-defining technologies.

When you can actually represent somebody like the ESA get in touch with them. Otherwise, uh, gtfo.