Pulse Daily — Wednesday, October 7, 2026

Let's start in Paris.

Mistral AI unveiled Mistral Large 4, officially nicknamed LeChonc, yesterday, October 6th, and the name is doing a lot of work.

This is a 1.05 trillion parameter, natively multimodal mixture of experts model with 49 billion parameters active on every single token.

Mistral trained it from scratch on 3,800 NVIDIA Grace Blackwell GPUs housed in its own European data centers, and the company is marketing it under the slogan, Forged in Europe, Built for AI Sovereignty.

The API preview went live immediately at $1.36 per million input tokens and $4.18 per million output tokens.

Full open waits, and this is the part the developer community is watching closely, are promised by October 27th, though the license terms haven't been published yet.

Mistral Large 3 shipped under Apache 2.0. Large 4's terms are still a question mark, and until those waits drop, developers can only benchmark a hosted preview, not audit or fine-tune the actual model.

And the benchmark picture is genuinely complicated, Amber.

Mistral is claiming this is the strongest open-weight model built outside China.

Independent data from artificial analysis scores the preview at 38 on its intelligence index, behind GLM 5.3 at 45 and KIMI K3 at 44.

Val's AI places it 32nd out of 44 models overall.

though it ranks first among openweight models on the legal agent benchmark from Harvey.

So there's a real gap between Mistral's framing and third-party scores.

And that gap won't fully resolve until October 27th, when anyone can download, probe, and test the actual weights.

The sovereignty pitch lands politically in Brussels, where EU digital autonomy policy has been a live debate for years.

But for practitioners, the license is the real story.

Will it be truly open? Or will Mistral introduce commercial restrictions the way some Chinese labs have? We'll know in three weeks.