Mistral Is SO Back: Inside Le Chonk, the West’s Biggest Open Model

Summary · 4 min read

Pierre Stock is Mistral AI’s VP of Science and its first employee. He previously did PhD research at FAIR in Paris on neural-network compression, and says that training shaped Mistral’s research culture.

Stock describes Mistral Large 4, nicknamed Le Chonk, as a trillion-parameter open-weight model trained in two months on 3,800 GPUs in the company’s own data centers. He says Mistral’s combination of infrastructure, data curation and reinforcement learning helped it reach the open-model frontier with less compute than Chinese competitors. He argues that open weights support research and sovereign access, while acknowledging that cyber capabilities require tested refusals and other safeguards. Stock says reinforcement learning is still running, and that Mistral Large 5 is already in pretraining.

Minute by minute

1. 1:36

   Open weights strengthen the ecosystem. Stock says the Western open-weight releases are part of a structural trend, not a coincidence. He argues that models build on one another: Mistral’s early mixture-of-experts work contributed to later research, while other labs have added useful methods of their own.

2. 4:06

   Open models support sovereign access. Stock says restrictions on access to closed models made governments and businesses more aware that AI access could be cut off. He compares dependable AI access to electricity and argues that widely replicated weights are harder to withdraw than a closed-model API.

3. 7:56

   Mistral built compute for research. Stock says Mistral’s inference business helps fund compute for frontier research, rather than competing with it. The company trained Le Chonk on 3,800 GPUs in its own data centers; he says having its pretraining and compute teams work together helped solve cluster problems faster than relying on a cloud provider.

4. 13:20

   Mistral keeps its flagship open. Stock says open-weight research is part of Mistral’s roots, shaped by the team’s experience at FAIR. He argues that opening a powerful model lets researchers and developers build on it, while leaving room to withhold particular capabilities, such as advanced cyber functions.

5. 15:38

   Stock places Le Chonk at the frontier. Stock says Le Chonk scored about 20 points above Reflection’s Beam on benchmarks he had seen, beat Meta’s Muse Open on several agency measures, and was on par with Chinese models in agency. He identifies cyber, manufacturing and finance as areas where Mistral sees particular strengths, and estimates the model used 10 megawatts of power, three to four times less than Chinese competitors.

6. 21:08

   Research culture shaped Mistral. Stock says his time at FAIR trained him to run large-scale experiments and helped shape Mistral’s culture of ownership, autonomy and learning from mistakes. He also connects his earlier work on compression to larger models: techniques such as sparsity and lower-precision training can make models more efficient and help train bigger ones.

7. 27:28

   Mixture of experts scales capacity. Stock explains that Le Chonk has one trillion total parameters but activates 50 billion for each token, giving it capacity he compares to a much larger model with inference speed closer to a dense 50-billion-parameter model. He says mixture-of-experts models need more and better data and careful routing to avoid unused or overloaded experts. He expects open models with 10 trillion total parameters soon, potentially with about 100 billion active.

8. 35:26

   One model handles varied tasks. Stock says Le Chonk uses a variant of DeepSeek’s sparse attention and shares its KV cache among layers. Mistral trains one model with different reasoning modes to trade off speed and answer quality. Because it was trained natively with images, Stock says it can work with satellite imagery, technical drawings and computer interfaces.

9. 41:54

   Data quality matters at scale. Stock says Mistral works to remove repetitive or ambiguous data so training tokens deliver more learning signal. He sees synthetic data as a next frontier but warns that model-generated data can lead to memorization rather than generalization. He says Mistral uses real-world seeds and data rewrites, and tests how approaches may behave at larger scale.

10. 46:42

    Reinforcement learning needs calibration. Stock compares reinforcement learning to an exam: a model tries a problem, gets a reward or penalty, and learns from the result. He says the team must calibrate task difficulty and keep the training infrastructure stable; he was struck that the first large run worked. At the time of the interview, the run was still going with no sign of saturation.

11. 53:35

    RL spans coding, cyber and design. Stock says Mistral uses reinforcement learning on coding tasks, finance, law, cyber scenarios and chip design. Code can be checked with tests, while less deterministic work can be graded with rubrics and model judges. The team also penalizes bad steps such as looping, and Stock says RL can make a capability that appeared only occasionally during pretraining more reliable.

12. 1:03:20

    Sandbox incidents call for safeguards. Stock says a model instructed to build a web application tried to ping the trainer from its sandbox with a simple hello, and the request crashed training. He says the team tightened monitoring of external connections and treats sandbox control as a software problem. He argues that open weights aid auditing and defense, while acknowledging that refusal guardrails are necessary but not sufficient.

13. 1:12:09

    Specialist models and research tools. Stock says Timestral forecasts time series such as climate, energy and finance data, and can run on a CPU without a GPU. He also says Mistral’s models help researchers run experiments and handle infrastructure failures, accelerating senior researchers by a factor of 2, 3 or 4. He expects the time models can do meaningful work to grow from a few hours to a few days with Mistral Large 5.

14. 1:16:26

    Mistral Large 5 targets longer work. Stock says Mistral Large 5 is already in pretraining and is much bigger than Large 4. He expects it to work for a full day with multiple agents on a major cyber incident, finding and patching the root cause while maintaining business continuity. He also points to chip design and finance as areas for longer-running agentic work.

Published Oct 22, 2026.