Removed, enjoy !

We’re announcing a new optimised model today! Mistral Large has top-tier reasoning capacities, is multi-lingual by design, has native function calling capacities and a 32k model. The pre-trained model has 81.2% accuracy on MMLU. Learn more on mistral.ai/news/mistral-larg…. Mistral Large is available on la Plateforme, as well as on Azure, following our commitment to bring frontier AI everywhere. We're eager to see what developers do with Mistral Large!

Self-qualifying oneself as heavyweight while shipping nothing of significance looks like hubris to me

We are announcing €600M in Series B funding for our first anniversary.  We are grateful to our new and existing investors for their continued confidence and support for our global expansion. This will accelerate our roadmap as we continue to bring frontier AI into everyone’s hands.

An over-enthusiastic employee of one of our early access customers leaked a quantised (and watermarked) version of an old model we trained and distributed quite openly. To quickly start working with a few selected customers, we retrained this model from Llama 2 the minute we got access to our entire cluster — the pretraining finished on the day of Mistral 7B release. We've made good progress since — stay tuned!

At @MistralAI we're releasing our very first model, the best 7B in town (outperforming Llama 13B on all metrics, and good at code), Apache 2.0. We believe in open models and we'll push them to the frontier mistral.ai/news/about-mistra… Very proud of the team !

A new model to hasten AI progress. 24B, 81% MMLU, no RL for now! We're super excited to see the latest development in international open-source AI (kudos to Deepseek!), and cannot wait to bring new contributions to it. We're renewing our commitment to using Apache licenses. AI brings joy: accelerate.

We have heard many extrapolations of Mistral AI’s position on the AI Act, so I’ll clarify. In its early form, the AI Act was a text about product safety. Product safety laws are beneficial to consumers. Poorly designed use of automated decision-making systems can cause significant damage in many areas. In healthcare, a diagnosis assistant based on a poorly trained prediction system poses risks to the patient. Product safety regulation should be proportional to the risk level of the use case: it is undesirable to regulate entertainment software in the same way as health applications. The original EU AI Act found a reasonable equilibrium in that respect. We firmly believe in hard laws for product safety matters; the many voluntary commitments we see today bear little value. This should remain the only focus of the AI Act. The EU AI Act now proposes to regulate “foundational models”, i.e. the engine behind some AI applications. We cannot regulate an engine devoid of usage. We don’t regulate the C language because one can use it to develop malware. Instead, we ban malware and strengthen network systems (we regulate usage). Foundational language models provide a higher level of abstraction than the C language for programming computer systems; nothing in their behaviour justifies a change in the regulatory framework. Enforcing AI product safety will naturally affect the way we develop foundational models. By requiring AI application providers to comply with specific rules, the regulator fosters healthy competition among foundation model providers. It incentivises them to develop models and tools (filters, affordances for aligning models to one's beliefs) that allow for the fast development of safe products. As a small company, we can bring innovation into this space — creating good models and designing appropriate control mechanisms for deploying AI applications is why we founded Mistral. Note that we will eventually supply AI products, and we will craft them for zealous product safety. With a regulation focusing on product safety, Europe would already have the most protective legislation globally for citizens and consumers. Any foundational model would be affected by second-order regulatory pressure as soon as they are exposed to consumers: to empower diagnostic assistants, entertaining chatbots, and knowledge explorers, foundational models should have controlled biases and outputs. Recent versions of the AI Act started to address ill-defined “systemic risks”. In essence, the computation of some linear transformations, based on a certain amount of calculation, is now considered dangerous. Discussions around that topic may occur, and we agree that they should accompany the progress of technology. At this stage, they are very philosophical – they anticipate exponential progress in the field, where physics (scaling laws!) predicts diminishing returns with scale and the need for new paradigms. Whatever the content of these discussions, they certainly do not pertain to regulation around product safety. Still, let’s assume they do and go down that path. The AI Act comes up with the worst taxonomy possible to address systemic risks. The current version has no set rules (beyond the term highly capable) to determine whether a model brings systemic risk and should face heavy or limited regulation. We have been arguing that the least absurd set of rules for determining the capabilities of a model is post-training evaluation (but again, applications should be the focus; it is unrealistic to cover all usages of an engine in a regulatory test), followed by compute threshold (model capabilities being loosely related to compute). In its current format, the EU AI Act establishes no decision criteria. For all its pitfalls, the US Executive Order bears at least the merit of clarity in relying on compute threshold. The intention of introducing a two-level regulation is virtuous. Its effect is catastrophic. As we understand it, introducing a threshold aims to create a free innovation space for small companies. Yet, it effectively solidifies the existence of two categories of companies: those with the right to scale, i.e., the incumbent that can afford to face heavy compliance requirements, and those that can’t because they lack an army of lawyers, i.e., the newcomers. This signals to everyone that only prominent existing actors can provide state-of-the-art solutions. Mechanistically, this is highly counterproductive to the rising European AI ecosystem.  To be clear, we are not interested in benefiting from threshold effects: we play in the main league, we don’t need geographical protection, and we simply want rules that do not give an unfair advantage to incumbents (that all happen to be non-European). Transparency around technology development benefits safety and should be encouraged. Finally, we have been vocal about the benefits of open-sourcing AI technology. This is the best way to subject it to the most rigorous scrutiny. Providing model weights to the community (or even better, developing models in the open end-to-end, which is not something we do yet) should be well regarded by regulators, as it allows for more interpretable and steerable applications. A large community of users can much more efficiently identify the flaws of open models that can propagate to AI applications than an in-house team of red-teamers. Open models can then be corrected, making AI applications safer. The Linux kernel is today deemed safe because millions of eyes have reviewed its code in its 32 years of existence. Tomorrow’s AI systems will be safe because we’ll collectively work on making them controllable. The only validated way of working collectively on software is open-source development. Long prose, back to building!

Announcing Mixtral 8x7B mistral.ai/news/mixtral-of-e… and our early developer platform mistral.ai/news/la-plateform…. Very proud of the team!

Introducing the world's best edge models. mistral.ai/news/ministraux/

Clarifying a couple of things since we’re reading creative interpretations of our latest announcements: - We’re still committed to leading open-weight models! We ask for a little patience, 1.5k H100s only got us that far. - We have a reselling agreement with Microsoft, that we’re very excited about. Alongside similar partnerships, it will accelerate our growth. - Microsoft invested in a small convertible note alongside many other companies, as a distribution partner. We are an independent European company with global ambitions, that part is not changing either. We’re seeing some interest for Le Chat and Mistral Large, on both la Plateforme and Azure, and we’ll be iterating fast!

We're back to school! Very proud of our team accomplishments, and honored to partner with @ASMLcompany in our next phase. We're very excited to push frontier AI capabilities in science and technology, with exciting releases ahead.

Congratulations, interesting how Github stars seem to correlate to superfluous parameters 😉

It may soon be a crime to compress public domain human knowledge into public domain matrices. We need to regulate the usage of AI in applications, not gradient descent

Mixtral is now powering Leo, Brave browser assistant!

Our new OCR model is available through our public API and as a self-deployable solution. In use on le Chat, and part of our specialist model family that is getting bigger !

Official now, very proud of the team! Apache 2.0 and instructed versions for your pleasure, available today on la Plateforme mistral.ai/news/mixtral-8x22…

The team is fast! It's been super exciting to see le Chat more and more widely adopted. It's an early product, and we can't wait to show you what's coming next. mistral.ai/en/news/all-new-l…

Our first shot at audio

Our work w/ @mblondel_ml 'Differentiable Dynamic Programming for Structured Prediction and Attention' was accepted at @icmlconf ! arxiv.org/abs/1802.03676 Sparsity and backprop in CRF-like inference layers using max-smoothing, application in text + time series (NER, NMT, DTW)

mistral.ai/news/devstral we’ve released an open-source model which is great at agentic coding tasks - by far Pareto optimal and a good prelude to what’s coming next

Complex matters slowly coming together — we actually got surprised ourselves

Today, we're announcing Mistral NeMo, a tiny multilingual model, 128k context length, trained with quantization awareness in collaboration with the NVIDIA research team.

As the Olympic season reaches Paris, we express our respect to ancient Greeks by releasing two new research models, MathΣtral and Codestral Mamba.

Triangle of happiness!

Expanding from a science company to a science and product company was no easy task, and that release is a very significant milestone in our journey. We're looking forward to how you'll use le Chat, now a slightly more mature animal

Reasoning with latency-optimized models is quite a UX game changer. Super proud of what the team has accomplished with this Magistral release! mistral.ai/news/magistral

This person has put a Mistral 7B model into a stuffed parrot and displays Chinchilla Equation (2) on his torso. Does anyone have his number? This seems a little unsafe.

As a small surprise, we’re also releasing le Chat Mistral, a front-end demonstration of what Mistral models can do. Learn more on mistral.ai/news/le-chat-mist…

Mistral is proud to provide the text LLM powering Unmute, the open-source voice AI from @kyutai_labs!

Welcome to the party

Our first hackathon! We’ll be in SF the week before to meet our users

With Codestral, our newest state-of-the-art code model, we are introducing the Mistral AI non-production license (MNPL). It allows developers to use our technology for non-commercial use and research. It ensures that every actor on the value chain builds successful businesses. mistral.ai/news/mistral-ai-n….

Moving forward with a new Small, Pixtral on le Chat and la Plateforme, reduced prices across the board, and an API free tier!

Day 1 delivery as always :)

Leaving the AI Safety Summit after some constructive discussions today and yesterday. I voiced how open-source was today the safest way to develop AI, putting this transformative technology under the highest level of scrutiny. With many others, we recalled the enormous opportunities that AI brings for better education, better healthcare, unlocking critical science problems, and making our jobs more rewarding. Transparency and fast information flow across different actors are needed to continue enabling these opportunities. Finally, we were many to stress how any new institutions measuring AI progress should be fully independent to avoid the pitfalls of regulatory capture. And it appears we have been heard! We must ensure that any such institution involves the entire world. More data is needed to understand the limitations of current models, and we look forward to engaging with the AI community to jointly agree on a scientific monitoring framework around new AI capabilities in the coming year.

At Mistral, we've grown aware that to create the best AI experience, one needs to co-design models and product interfaces. Pixtral was trained with high-impact front-end applications in mind and is a good example of that.

It’s removed, we missed it in our final review — no joke of ours, just a lot of materials to get right !

"Online Sinkhorn" refines an Optimal Transport distance between two continuous distributions, based on a stream of samples. Stochastic approximation again ! Joint work with @gabrielpeyre, to be presented (oral) at #NeurIPS2020. arxiv.org/abs/2003.01415 github.com/arthurmensch/onli…

Flamingo does feel slightly conscious these days 🦩

Apache 2.0 indeed

Local Codestral generating #scikitlearn and Keras code to run a medical imaging prediction on the fly. I like that :)

Self deployable fully packaged frontier AI. If you’re a country you may want to look here instead, we can help!

Very proud of the team, a first step towards making model customisation much simpler.

Very excited to be bringing our models to Snowflake customers as part of this multi-year partnership. LLMs become all the more interesting when contextualised on data, and we’re eager to see developers create powerful applications combining Mistral models with the Data Cloud.

Very happy to partner with @awscloud to expose Mistral models on Amazon Bedrock, as we continue to bring our technology to every developer.

Mistral models will very soon be available as a service on Azure, thank you @satyanadella! We bring our technology where developers build.

Mistral Medium 2 was Miqu by the way

Codestral 25.01 is not only on top of the Copilot Arena leaderboard, it's also 2x faster than the first Codestral -- that matters a lot for code completion

3 notebooks on @PyTorch : from optimization (autodiff basics) to learning (a 2-parameter MLP) to deep-learning (Fashion Mnist + learning rate tricks). Thanks @ogrisel @CharlesOllion for this collaboration ! github.com/m2dsupsdlclass/le…

Notice periods are our biggest pain by far

Welcome Marie!

The safest way to make useful AI while mitigating misuse is to work on improving it in the open, with the highest level of scrutiny — as always in software. That’s why open-source is the one and only path toward AI safety.

We are very excited to partner with Harvey to help build domain-specific models! Our platform's deployment flexibility and high customisation capabilities will help Harvey address the highly regulated legal industry. harvey.ai/blog/mistral-annou…

Totally thrilled to be alongside @GuillaumeLample and @tlacroix6 to create Mistral AI. A lot of work ahead of us!

We shipped higher level apis for agent orchestration - MCP compatible, all server-side logic deployable privately.

A first step towards faster and stronger differentiated models for your use cases. We've rethought fine-tuning to make it super easy to use, both at training and inference.

AI needs to be connected to the physical world, proud to be supporting !

Largement sur le Pareto

Mistral 7B is now in prod, nice work @perplexity_ai !

"Geometric losses for distributional learning" w/ @mblondel_ml and @gabrielpeyre accepted @icmlconf. We derive a geometric softmax with reg. optimal transport and Fenchel duality. Accounts for cost bw/ classes, outputs discrete/*continuous* distributions. bit.ly/2HkBDKq

To seriously compare two methods in deep learning (say, GANs) , you're looking at 2 methods * 5 random seeds * 5 learning rates * 2 day = 100 days of GPU usage, and that's a minimum. This is often beyond academic labs capacity, how do you cope?

Still committing mistral.ai/ ;)

New steps towards faster model customisation and application building

(sorry for self-promotion) I will be presenting our work on a new geometric softmax with continuous output, based on OT and Fenchel duality, tonight 6:30 @icmlconf, poster #179. Joint work with @mblondel_ml and @gabrielpeyre

New steps toward completing our AI platform - proud of the team!

Merci pour votre confiance @SebLecornu, construisons ensemble l'IA de Défense 🇫🇷

We need to hammer this home: the soon-to-become AGI of many is mere statistical knowledge compression. It’s great, it can help solving millions of problems in healthcare, in education, it can make the jobs of everybody more creativity-oriented.

Very proud to be partnering with @Capgemini to help enterprises adopt our technology!

Very proud of this massive scale partnership. We’re investing in infrastructure to bring AI strategic autonomy to the entire world

We purposely made it great at optimal transport as you may have guessed !

We're presenting RETRO at 4:15pm @icmlconf with @borgeaud_s, and later today at the poster session. Add a retrieval DB to divide your model size by 10, don't miss out!

Excited to partner with Cloudflare on low-latency generative AI!

At NeurIPS this week, reach out if you want to discuss our work around LLM at @DeepMind (Chinchilla, Retro, Flamingo), and if you're interested in working with us!

We need less religion and more science to make AI safe and useful

Happy to present our work on games @icmlconf: Gradient extrapolation with alternated player update speeds-up equilibrium finding in convex games/GANs ! Joint work with S. Jelassi, C. Domingo, D. Scieur, @joanbruna @NYUDataScience. arxiv.org/pdf/1905.12363.pdf github.com/arthurmensch/dseg

Thank you for your trust Chuck, we're extremely excited to be working with Cisco

I had the great pleasure to talk about sovereignty in AI with Jensen and @AnjneyMidha. We explain why every nation state needs an AI strategy and what matters for it to be successful. Full link below

We’re bringing le Chat and our AI Platform to Dell’s hardware for on-prem deployment dell.com/en-us/blog/bringing…

Can this change? Yes, although not without a paradigm change in the technology, and that’s why it’s great that we talk about risks. But right now the true risk is to mechanically leave the development of AI to 2 or 3 large corporations.

Hello, thanks for this study! It seems you have been using a third-party model github.com/neulab/gemini-ben…… based on Mixtral base model. You'll probably get better results by working with our instructed version (mistral-small on our API). Let us know if we can help!

DX improvements on la Plateforme!

Answered some hard questions during our trip to SF. Thanks @jacobeffron and @jordan_segall for hosting!

But it’s not that hard to replicate. Anyone with 100M and the will to do it can create a rather good model from anywhere on earth. Good actor, or bad actor.

We present our work on multi-task learning in fMRI at poster #151 #NIPS2017 this afternoon with @GaelVaroquaux @julienmairal @danilobzdok, feel free to come and discuss ! arxiv.org/abs/1710.11438

Very happy to see the release of these works ! In particular, we push semi-parametric language models (a AR Transformer and a nearest-neighbor database) to an unprecedented scale, and obtain continuous improvements with DB size. The semi-parametric route holds many promises !

Super excited by this partnership with @SnowflakeDB, with specialist models!

The good news is, for all of the illegal usage of AI we can imagine (misinformation and knowledge search), AI currently does nothing to lift the actual bottleneck standing in the way (distribution of information, actual execution of what the LLM recommends doing, respectively)

Great to see the release of Llama II, open-source LLMs are making good progress! Still a lot of room to improve OS models positioning on the efficiency/performance front — so that they eventually catch up with proprietary solutions. An interesting challenge 😇

Canevas is a French word!

"It directly follows from Koshi-Schwartz inequality", welcome to a subtitled #NeurIPS2020 conference that promises to be fun

Our paper #nips2017: Multi-layer classif° and dropout regul° permit transfer learning between fMRI datasets. goo.gl/WrnWqb

solved :)

my number one priority

I will be presenting this work at @icmlconf on Thursday morning, and at poster #48. Please come and discuss :-) (alpha) code online github.com/arthurmensch/didy…

Our paper on faster stochastic matrix factorization to appear in IEEE TSP 😃 Convergence proofs, 10x speed-ups, code. goo.gl/F5Jgpe

My slides #icml2016: online learning + random subsampling for fast matrix factorization. Thanks to the audience ! amensch.fr/docs/presentation…

What an abusive subtitle system @NeurIPSConf 😇 I think there is a slight bias against the French accent, and towards French in general: the optimal transport distance (a national pride) was systematically dubbed the American transport distance !

Large language models are too big: very happy to share this work which is a welcome finding for computational sobriety! If you're into training big models, consider a dataset collection specialist for your next hire, more text tokens are needed ;)