Mistral AI Releases Forge

by pember • 4 months ago

733 192 4 months ago

roxolotl• 4 months ago

Mistral has been releasing some cool stuff. Definitively behind on frontier models but they are working a different angle. Was just talking at work about how hard model training is for a small company so we’d probably never do it. But with tools like this, and the new unsloth release, training feels more in reach.

csunoser• 4 months ago

Huh. I initially thought this is just another finetuning end point. But apparently they are partnering up with customers on the pretraining side as well. But RL as well? Jeez RL env are really hard to get right. Best wishes I guess.

mark_l_watson• 4 months ago

I am rooting for Mistral with their different approach: not really competing on the largest and advanced models, instead doing custom engineering for customers and generally serving the needs of EU customers.

w4yai• 4 months ago

Go Mistral !

jerrygoyal• 4 months ago

their ocr model is goated

stavros2• 4 months ago

Better than Qwen? I guess the best overall is Gemini, right?

SyneRyder2• 4 months ago

Did they make significant improvements in OCR 3? The quality I was getting from Mistral OCR 2 was nowhere near as good as what I could get from just sending the same files to Claude Sonnet via an API call.

I have been finding Voxtral useful though.

oakpond• 4 months ago

probably yes. considering that even some of their non-ocr models can recognize my shitty handwritten math

doctorpangloss• 4 months ago

first, there was .ai

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

but what about ai.eu

fnord123• 4 months ago

but what about ai.eu

oh, .. why?

ChrisGreenHeur• 4 months ago

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.

nicman23• 4 months ago

also offering support for local deployments

bsjshshsb• 4 months ago

Id training or FT > context? Anyone have experience.

Is it possible to retrain daily or hourly as info changes?

rorylawless• 4 months ago

The fine tuning endpoint is deprecated according to the API docs. Is this the replacement?

https://docs.mistral.ai/api/endpoint/deprecated/fine-tuning

aavci2• 4 months ago

Interesting to see. I thought they were promoting fine tuning

dmix• 4 months ago

This is definitely the smart path for making $$ in AI. I noticed MongoDB is also going into this market with https://www.voyageai.com/ targeting business RAG applications and offering consulting for company-specific models.

andai2• 4 months ago

They mention pretraining too, which surprises me. I thought that was prohibitively expensive?

It's feasible for small models but, I thought small models were not reliable for factual information?

simsla• 4 months ago

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.

ryeguy_24• 4 months ago

How many proprietary use cases truly need pre-training or even fine-tuning as opposed to RAG approach? And at what point does it make sense to pre-train/fine tune? Curious.

baby3• 4 months ago

[flagged]

charcircuit3• 4 months ago

Using tools and skills to retrieve data or files is anything but dead.

CharlesW2• 4 months ago

[flagged]

bigyabai• 4 months ago

In what, X's hype circles? Embeddings are used in production constantly.

loeg• 4 months ago

Is it??

Shitty-kitty• 4 months ago

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.

mirekrusin2• 4 months ago

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.

troyvit2• 4 months ago

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.

aavci2• 4 months ago

How does this compare to fine tuning?

Otterly99• 4 months ago

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

hermit_dev2• 4 months ago

The future of AI is specialization, not just achieving benevolent knowledge as fast as we can at the expense of everything and everyone along the way. I appreciate and applaud this approach. I am looking into a similar product myself. Good stuff.

reverius42• 4 months ago

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.

holoduke• 4 months ago

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.

hermit_dev2• 4 months ago

To be more specific, I think the future is local and specialized. IBM among others thought the same way with their giant mainframe centralized computers and the original way people would utilize software in the 70s. It's an interesting parallel to today's cloud if you think about it. It's just not scalable from a resource (hardware), energy, and cost perspective. I think we're living a unique time, but it's going to change. Without continued massive funding and a pivot to sustainable, things will (and should) change.

Don't get me wrong, general intelligence will always be important and should be a part of specialist models to a degree for understanding, but it doesn't make sense to use an 800B+ parameter model to help write an email or do research on company trends. Hell, look at what China has been able to do. Qwen 3.5 9B, exceeds Claude 3.5 Haiku and nears Sonnet 3.5 levels. The 27B variation of Qwen 3.5 is superior to both in many ways and even rivals newer models. There is obviously an inherit lag behind, but we will gradually see a shift as these models become more capable.

Right now we are chasing 1-2% improvements at the cost of billions. Local are already absurdly capable (more and more by the day - same with cloud ofcourse) and smarter than most people in specific areas. To do most jobs, can we honestly say it requires a PhD or higher level understanding to perform? We're chasing something that is becoming more and more not needed from a general day to day perspective. AGI is outstanding, but not practical (at least today). I think we'll get there anyway at our current trajectory (though dangerous), but I suspect things will shift.

upghost2• 4 months ago

Pre-training allows organizations to build domain-aware models by learning from large internal datasets.
Post-training methods allow teams to refine model behavior for specific tasks and environments.

How do you suppose this works? They say "pretraining" but I'm certain that the amount of clean data available in proper dataset format is not nearly enough to make a "foundation model". Do you suppose what they are calling "pretraining" is actually SFT and then "post-training" is ... more SFT?

There's no way they mean "start from scratch". Maybe they do something like generate a heckin bunch of synthetic data seeded from company data using one of their SOA models -- which is basically equivalent to low resolution distillation, I would imagine. Hmm.

stingraycharles• 4 months ago

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.

anon373839• 4 months ago

I think they are referring to “continued pretraining”.

mirekrusin2• 4 months ago

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

gunalx• 4 months ago

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

lelanthran7• 4 months ago

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).

qntty• 4 months ago

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.

losvedir2• 4 months ago

Er, then what is the "already trained" model? I thought pre-training was the gradient descent through the internet part of building foundational models.

upghost2• 4 months ago

Yeah, this checks out. I wonder what they are doing to prevent semantic collapse. Also, I wonder if the base model would already be instruct and RLHF tuned or only pre-trained. Trying to do additional training without semantic collapse in a way that is meaningful would be interesting to understand. Presumably they are using adapters but I've never had much luck in stacking adapters.

i.e.:

1. Do I start with an RLHF tuned model, "pretrain" on top of that (with adapter or by freezing weights?), then SFT on top of that (stack another adapter, or add layer(s) and freeze weights?) (and where did I get the dataset? synthetic extraction from corpus?), then RL (adapter, add layer(s) and freeze?)

2. or do I start at SF tuned model, ...

3. or do I start at raw pre-trained model, ...

Would love to know what the matrix used was.

ogou• 4 months ago

Don't sleep on Mistral. Highly underrated as a general service LLM. Cheaper, too. Their emphasis on bespoke modelling over generalized megaliths will pay off. There are all kinds of specialized datasets and restricted access stores that can benefit from their approach. Especially in highly regulated EU.

Not everyone is obsessed with code generation. There is a whole world out there.

haraldooo• 4 months ago

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

isodev• 4 months ago

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.

srivmo• 4 months ago

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?

lelanthran7• 4 months ago

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?

Think of it as a base model (the megalith) which then has the weights adjusted towards a specific use-case (SAP, for example).

wodenokoto• 4 months ago

ChatGPT is already a small agent that receives your message and decides which agent needs to respond. Within those, agents can have sub agents (like when it does research).

When generating images most services will have a small agent that rewrites your request and hands it off to the generative image model.

So from the treasure trove point of view, optimized agents have their place. From companies building pipelines, they also have their place.

Stromgren• 4 months ago

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.

umeridrisi• 4 months ago

Is this the best Grok alternative?

spiderfarmer7• 4 months ago

Any model is.

spiderfarmer7• 4 months ago

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

jstummbillig2• 4 months ago

Why would that be? Most big EU companies use ms teams or google workspace, for example.

lelanthran7• 4 months ago

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.

Lapel27423• 4 months ago

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

My 2ct: Currently the moat may be that they are not US-American which is not reproducible by any of the US alternatives.

erispoe• 4 months ago

Except the evidence today rather points to SOTA model + harness than fine tuned models.

butILoveLife5• 4 months ago

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.

danelski2• 4 months ago

Its bottom floor for LLMs.

What? That's just demonstrably false. The market doesn't consist of 5 providers.

Forgeties79• 4 months ago

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

Bombthecat2• 4 months ago

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

rldjbpin• 4 months ago

Not everyone is obsessed with code generation.

They (still) are. https://news.ycombinator.com/item?id=47404796

supernes• 4 months ago

Code agents are becoming the primary users of developer tools, so we built Forge for them first, not

... for humans.

dash2• 4 months ago

I think it’s interesting what this approach suggests about who will profit from AI. I’m sceptical that having huge numbers of GPUs is a moat. After all, real humans – even geniuses – are trained on much much less data than the whole Internet. But proprietary and specialised data could very well be a moat. It’s hard to train a scientist/lawyer/analyst without reading a lot of science/law/finance. Companies’ proprietary data might encode a great deal of irreplaceable knowledge. Seems as if Mistral is taking this bet.

copirate• 4 months ago

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.

speedgoose• 4 months ago

I was enthusiastic but it’s "contact us" priced for now. I was expecting a classic cloud LLM forge with a public pricing.

0x457• 4 months ago

This feels like something more hands-on rather than "here some tools to do this"

zby4• 4 months ago

My bet is that the solution to continuous learning is with external storage. There is a lot of talk about context engineering - but I have not seen anyone taking context as the main bottleneck and building a system around that. This would show that even context engineering is kind of wrong term - because context does not enter the llm in some mysterious way - it goes through prompt and the whole model of passing chat history back and forth is not the most efficient way of using the prompt limitation.

Centigonal• 4 months ago

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

zby4• 4 months ago

A knowledge base - something where the LLM knows how to find the knowledge it needs for a given task. I am working on this idea in https://zby.github.io/commonplace/

ithkuil• 4 months ago

A form of context engineering

mhl47• 4 months ago

"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.

zby4• 4 months ago

The key to make Copilot useful is to take the limited context problem seriously enough. There are many dimensions to it: https://zby.github.io/commonplace/notes/context-efficiency-i... and it should be the starting point for designing the systems that extensively use llms.

whatever1• 4 months ago

I thought that for pretraining to work and reasoning to emerge you need internet scale data. How can forge achieve it with just internal company data (unless the said company is AT&T or something) ?

jbverschoor• 4 months ago

ASML and ESA as clients means something. I dont expect to see the first name somewhere else on the logo list

kioleanu4• 4 months ago

I like Mistral, it hits the exact sweet spot between cost and my data staying in the EU, withouth a significant drop in quality, but man are their model naming conventions confusing af. They mention they have a model called Devstral 2, which is neither Codestral nor Devestral. I want to use it, but the api only lists devstral-2512, devstral-latest, devstral-medium-latest, devstral-medium-2507, devstral-small, devstral-small-2507.

I think, devstral-latest should be it, no? So I write to support and get an answer 12 hours later that says oh, no, devstral 2 is definetely called devstral 2 and then a page of instructions on how to set it up in Intellij... generated with AI. The screens it is refering to don't exist and never did.

Manfred• 4 months ago

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?

kioleanu4• 4 months ago

no, the key is actually universal, you can't choose a specific product

newswasboring2• 4 months ago

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.

kioleanu4• 4 months ago

you might be correct. for example, they have an intellij plugin that allows integration without the AI Assistant, but it is only available for Enterprise customers

soco2• 4 months ago

But wouldn't winning devs be a neat helping point in winning b2b contacts? Or they think golf courts are enough for success? Okay they might be right here, but still they make it so confusing for no obvious reason.

IanCal• 4 months ago

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

kioleanu4• 4 months ago

wow, thank you, this is great. I was thinking they should have a page like this, but I couldn't find myself.

butILoveLife5• 4 months ago

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.

Aldipower• 4 months ago

I cannot keep up with their products, model names and releases. What is what for? Their marketing texts do not make sense for me. Is there a nice overview somewhere?

I am a simple stupid Le Chat user with a small mind and the Tredict MCP Server connected to it (to Le Chat, not my mind), which works ok-ish. :-)

krinne• 4 months ago

I wasnt able to find a way to access this - is this something accessible only to enterprises ?

Would love to take it for a spin, if that is even possible.

thecopy• 4 months ago

Looks interesting. But how to explore or test or use? The product page (https://mistral.ai/products/forge) also does not contain anything useful. Just "Contact us"

Dissapointing.

burgerquizz• 4 months ago

can i use mistral to read my source code and teach it so i don't need to inject the whole doc every single time and consume token every single time?

spacesh1psoda• 4 months ago

Go EU!

jcmartinezdev• 4 months ago

Mistral is doing some really great stuff lately. Sure, it's hard to compete with OpenAI and Anthropic and their models, but they are taking up some interesting takes and designing their product in unique ways.

I like a lot what they are doing and I'll be watching them a lot more closely. I'd love to work for them btw!

apexalpha• 4 months ago

This looks good but how much money are we talking here? Are we 'retraining' an entire model but adding enterprise data to the public data set?

tho23i42342397• 4 months ago

Interesting. Does this actually scale though ? I've never seen enterprises which have "internal knowledge" in proper readable form - it's often in code, and more importantly in people who wrote them.

I recall that even at Google - with its own search engine and so on - the best way to understand anything was to read code or to reach out to those who wrote them. I don't know how it works in places that work with the "real world" like ASML.

Often the issue is not even about documentation - it's just that it's extremely hard to include all the nuances in text and still have it be readable (code-documentation comes to mind).

Interestingly, I strongly feel that this also where LLMs (and some of our more textually-obsessed academics) fail.

bob0012• 4 months ago

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.

dragochat• 4 months ago

where sample notebook/script? where github? where signup?

...learn a thing or two from NVIDIA or gtfo

troyvit2• 4 months ago

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.

losvedir2• 4 months ago

Forge enables enterprises to build models that internalize their domain knowledge. Organizations can train models on large volumes of internal documentation, codebases, structured data, and operational records. During training, the model learns the vocabulary, reasoning patterns, and constraints that define that environment.

I'm probably really out of date at this point, but my impression was that fine tuning never really worked that well for knowledge acquisition, and that don't variety of RAG is the way to go here. Fine tuning can affect the "voice", but not really the knowledge.

mikodin• 4 months ago

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

Havoc• 4 months ago

Good for them. Really hope they find market fit

alansaber• 4 months ago

I find the mistral "middle" between small LMs /1T LMs compelling. Models that are sufficiently big to be performant but specialised for domains and tasks- this is what I assumed we'd always head towards.

vincentbusch• 4 months ago

lol the AI-generated support reply about their own AI model is peak 2026

the naming mess is wild though. i ran into similar confusion trying to set up mistral for a side project — ended up just guessing which endpoint was the right one

Mistral AI Releases Forge

Related Stories