Pierre Stock

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

[Arthur Mensch @arthurmensch]

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

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

🎙️ 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 !

[Arthur Mensch @arthurmensch]

We've just released a comprehensive suite of state-of-the-art voice models https://t.co/Ue1S72c9HW. Voice will be the most important interface to AI systems, and we deploy voice models physically close to reasoning models

[Mistral AI @MistralAI]

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

[Surge AI @HelloSurgeAI]

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 🇫🇷

[Arthur Mensch @arthurmensch]

We’re building AGI from this train station

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

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

[Albert Jiang @AlbertQJiang]

Some things to say about this model:

1. It's a good model in the actual usage sense. I've understood so much about Lacan and Merleau-Ponty in the last two days
2. I spent a non-trivial amount of time to make it good, and then to make it servable
3. We had a tattoo bet about its DeepSWE performance being over 60 (a capability I am responsible for) so I welcome tattoo parlour recommendations to send my colleagues to
4. I marvelled at how good some of my colleagues are at neural surgery under stress
5. An even chonkier boi is munching on trillions of tokens at this very moment

https://t.co/fFkJZZZs4y

[Mistral AI @MistralAI]

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.

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

Today, we are launching a preview of our new model, Mistral Large 4 (ML4), aka le Chonk 🐈.
ML4 is a 1T-parameter model with 49B active parameters, trained natively with multimodal capabilities.

It is at the frontier of open models, and by far the strongest open-weight model from the US or Europe.
The RL run behind this preview is still in flight and shows no sign of saturation -- we will release a final version before the end of the month along with the weights of the model.
🧵
1/n

[Pierre Stock @PierreStock]

Alors, peut-etre.

[Arthur Mensch @arthurmensch]

Excited

[Arthur Mensch @arthurmensch]

Excited

[Mistral AI @MistralAI]

Today, we are announcing a partnership with @mozilla to bring privacy, control and choice to people using AI to browse online. 🦊🐈

https://t.co/Bsd6N4jNdV

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

We have just raised 3B€ to accelerate our R&D. We are scaling up our training and inference compute to build future generations of our models, and expanding our Science team across all our offices -- Paris, Palo Alto, SF, London, Zurich, and Warsaw. Join us !

[Arena.ai @arena]

The performance gap between proprietary and open source models has reached its narrowest point yet.

In Code Arena: WebDev, the frontier gap between proprietary and open has compressed from ~150 pts since late 2025, to just ~10 pts today.

Muse Spark 1.2 is expected to release their weights. Stay tuned to see how that and other open models like Glimmer impact the frontier. Official scores coming soon.

[Arthur Mensch @arthurmensch]

It's actually le gros chaton

[Pierre Stock @PierreStock]

Very excited to release Voxtral 2!
- Voxtral Realtime: a SOTA real-time transcription model under an Apache 2 license, with latency configurable to sub-200 ms
- Voxtral Mini Transcribe 2: a SOTA transcription model with diarization, word-level timestamps, and context biasing

[Guillaume Lample @ NeurIPS 2024 @GuillaumeLample]

Mistral robotics team is hiring. Join us!

[Mistral AI @MistralAI]

Introducing the Devstral 2 coding model family. Two sizes, both open source.
Also, meet Mistral Vibe, a native CLI, enabling end-to-end automation.
🧵