
That one took some groundwork. Trained and served on our own compute, and RL shows no sign of saturation

VP Science @mistralai. Ex RS @metaAI | PhD @ENSdeLyon | MVA @ENS_ParisSaclay

That one took some groundwork. Trained and served on our own compute, and RL shows no sign of saturation

Mistral VP of Science @PierreStock says they have efficiency in their DNA and people should give them more money: "Now, it's 300 people, and [Le Chonk] we managed to train it on only 10 megawatts of compute from scratch. 10 megawatts of compute is roughly 4,000 GB300s, to give you an order of magnitude." "That's super low. It's a fraction of what every competitor uses, and we trained it on our own compute." "It's pretty amazing, and we really got efficiency in our DNA. But now, the thing I want to tell everyone is that it means that you should give us more money because the yield is there....Please give us 1 gigawatt." "We've now spent one year building the reinforcement learning machine in training… [Le Chonk] is the best open-weight model in the world on cyber defense, on coding, and on coding agents."
ML4 is now open-weight! Frontier capabilities on code agents & cyber, way more efficient and compact than its Preview version three weeks ago (!) and cherry on the cake: technical report telling its epic backstory. Check it out: huggingface.co/mistralai/Mis…

Many positive things came out of this conversation with @pierrestock Mistral’s model Le Chonk is the best open-weight model in the world on cyber defense & coding, it’s one of the most efficient models in the world and they’ll be very powerful with 1 gigawatt of compute because of this efficiency. I've never been more bullish on this company.
Mistral VP of Science @PierreStock says they have efficiency in their DNA and people should give them more money: "Now, it's 300 people, and [Le Chonk] we managed to train it on only 10 megawatts of compute from scratch. 10 megawatts of compute is roughly 4,000 GB300s, to give you an order of magnitude." "That's super low. It's a fraction of what every competitor uses, and we trained it on our own compute." "It's pretty amazing, and we really got efficiency in our DNA. But now, the thing I want to tell everyone is that it means that you should give us more money because the yield is there....Please give us 1 gigawatt." "We've now spent one year building the reinforcement learning machine in training… [Le Chonk] is the best open-weight model in the world on cyber defense, on coding, and on coding agents."


ML4 is now open-weight! Frontier capabilities on code agents & cyber, way more efficient and compact than its Preview version three weeks ago (!) and cherry on the cake: technical report telling its epic backstory. Check it out: huggingface.co/mistralai/Mis…

Impressed by the new @MistralAI large-4 update! The model update is much better at coding and performs as the 2nd best OSS model on our coding benchmark

Good stuff, chonky stuff!
We just released an updated version of le Chonk (ML4) on the API and on HuggingFace. It is simply the best OSS model for code! Le Chonk pushes the OSS Pareto frontier for the AA coding agent index: tied with GLM5.3 using close to half of the tokens. It is the best OSS model on FrontierCode 2.0, which came out earlier this week, and on Terminal-Bench 4.0! 🐈

We sat down with @MistralAI's Pierre Stock and Sagar Vaze from Mistral on launch day of Mistral Large 4. Open weights. Trained in Mistral’s own data centers. Excels at cyber. Full conversation on cyber, training, evals and sandboxing: youtu.be/FczbSauVvjw

Today, we release the weights of a new and improved Le Chonk, advancing the open frontier for agentic coding and cybersecurity. 💻 Built to build: Le Chonk ranks #1 on @cognition’s FrontierCode among open models, and sets a new benchmark for Pareto efficiency on @artificialanalysis. 🛡 Built to defend: State-of-the-art in cyberdefense, put through the paces by the world's leading cyber vendors, including @CrowdStrike, @cloudflare, @rubrik, @zscaler, and others. Built for sovereignty: Make it yours. Download the weights on @huggingface , customize it with Mistral Forge, and deploy it wherever you need it. Read more from our team and customers 👇:

This Thursday, we are hosting events at our SF, NYC, and London offices to celebrate the launch of ML4 ! We're growing our teams across all three locations, as well as in Paris, Warsaw, Zurich, and Montreal, where we recently opened a new office. If you believe the future of AI should be open, that scientific breakthroughs should be shared, and that everyone should benefit from frontier intelligence, come build that future with us !
Le Chonk is out of the bag 🐈 We're throwing Chonky Parties in our SF, NYC and London offices to celebrate the launch! You can expect some Champagne 🍾, food, good music, and to meet the team behind our latest release. 📍 SF: luma.com/viqhwgcw 📍 London: luma.com/xvuvl3mc 📍 NYC: luma.com/1qk99m3e Please sign up! Registration is mandatory.

🎙️ Introducing Voxtral 3 ! A new family of SOTA speech models: ASR: Mini & Small for batch transcription, plus real-time streaming TTS: expressive, multilingual speech, batch + streaming Voice agent SDK: low-latency conversations + tool calling Available via API or on your own infrastructure !
Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵

We've just released a comprehensive suite of state-of-the-art voice models mistral.ai/news/voxtral-3/. Voice will be the most important interface to AI systems, and we deploy voice models physically close to reasoning models

Introducing Voxtral 3: state-of-the-art Transcribe and Speech models powering Mistral’s new voice agents API. 🧵

For their Mistral Large 4 release, @MistralAI ran a blind Surge human eval, with our expert software engineers evaluating five frontier models on coding quality. Le Chonk ranked #1 among open-weight models and #2 overall, behind only Opus 5. Unit tests ask: does it work? Human code review asks: would you merge it? Our evals measures the difference. Working code is the floor; professional taste and judgment are what make it shippable. Congrats to the Mistral team on le Chonk and their big release 🇫🇷
Meet Mistral Large 4, aka Le Chonk. • 1T parameters, natively multimodal. 49B active. It is the best open weights model from US or Europe on aggregated benchmarks. • State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding. • Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure. • Available to all via API today. Working with cybersecurity partners privately. Open weights release end of October.

We’re building AGI from this train station
ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise. We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster. We expect large and rapid improvements in the weeks and months ahead. 2/n

ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise. We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster. We expect large and rapid improvements in the weeks and months ahead. 2/n