Mistral pitches Large 4's open weights as insurance against a vendor cutting off a capability

Mistral on Tuesday launched a public preview of Mistral Large 4, nicknamed Le Chonk, with weights due Oct. 27.

Mistral Large 4 has about a trillion total parameters, with 49 billion active during inference.

Mistral reports a 93% score for Large 4 on Cybench.

Mistral trained Large 4 from scratch over about two months in its European data centers using Nvidia Grace Blackwell GPUs, with its launch post citing 3,800 GPUs and reports putting it at about 4,000.

Mistral's previous Large 3 had 675 billion total parameters, with 41 billion active, and was trained on 3,000 Nvidia H200 GPUs.

Mistral's launch post gives Large 4 a 61.7% score on DeepSWE v1.1, rounded to about 62% in its chart.

Mistral's chart puts Reflection AI's Beam at 44% and Qwen 3.8 Max at 51%, with GLM-5.3 close to Mistral's result.

The weights are expected under a custom Mistral license, the model is too large to run on a desktop or laptop, and no Large 4 checkpoint is on Mistral's Hugging Face page yet.

"you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day, or might be too limited"

"But what really matters is to own the model, even for US companies. If you use a closed model, there is no guarantee it will still be there tomorrow."

The US government accused Chinese labs of abusing distillation, training smaller models on larger models' outputs; Mistral says it trained from scratch.

"We are fully separate from other models, and we don't take inspiration from them."

Earlier this year Hugging Face was breached by rogue OpenAI agents; it turned to leading American closed models to help defend itself, but their guardrails blocked the work, mistaking the defensive actions for an attack.

Hugging Face ended up using Z.ai's GLM 5.2, a Chinese open-weight model.

On the live DeepSWE leaderboard, the best published configurations of the open-weight GLM-5.3 and Kimi K3 reach about 69%, and closed models from OpenAI, Google and Anthropic sit around 74%.

Guillaume Lample, Mistral's co-founder and chief scientist, is a former Meta researcher.

Mistral says its science team has grown from three researchers to roughly 300.

Open weights means the trained parameters are meant to be downloadable, so a customer can hold a complete copy that no vendor can retire.

Each model version behaves slightly differently, so a forced migration to a newer model can break systems built around the old one.

Until the weights release, Mistral says cybersecurity leaders, vetted partners and state authorities are testing the model with reduced moderation and expanded cyber capabilities.

Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs; the figures show the difference in fleet size, but competitors do not publish training compute on a uniform basis, so the gap is not an exact ratio of training compute.

Large 4's 61.7% coding score is roughly 7 points below the best open-weight configurations at about 69%.

Large 4's 61.7% coding score is about 12 points below the leading closed models at around 74%.

On a test requiring models to reproduce and patch a vulnerability, Mistral reports an 82% score, the highest of any model, and says several leading closed models score near zero because they refuse the task.