Mistral Large 4 Launches: 1T Parameters, Trained in Europe, Open Weights Promised

Mistral released Large 4 on October 6, and the release is a statement on every axis at once: 1 trillion total parameters with about 49 billion active, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs, served from Mistral’s own European datacenters, priced at $1.36/$4.18 per million tokens, and officially nicknamed “Le Chonk.” Open weights are promised “by the end of the month.” The release also marks the first milestone of a 3 billion euro Series D that Mistral claims is the largest European tech equity round ever, and it arrives as the most-discussed launch on Hacker News, with two simultaneous front-page posts.

The Benchmark Profile: Broad Frontier, One Notable Outlier

The numbers describe a genuinely frontier model rather than a regional champion. DeepSWE v1.1 at 61.7%, AutomationBench at 59.9% across 657 workflows, an AA-Briefcase Elo of 1,393, and a blind human eval via Surge AI ranking it second of five models behind Claude Opus 5 and ahead of Kimi K3 and GLM-5.3. Trained on 160-plus languages including every official EU language, it beat GPT-6 Astra on finance and legal tasks in vals.ai’s third-party evaluation and edged it on vision (42% versus 41% on Dense 200). The outlier is cyber: 82% on one AA Cyber Index test, which Mistral notes is the highest of any model, precisely because Claude Opus 5.5 and GPT-6 Astra score near zero on it due to refusals. That is the same philosophical divide as Google’s safety-gated Argon rollout, taken to its conclusion: Mistral ships the capable model and refuses less.

The Pricing War Gets a European Front

At $1.36/$4.18 per million tokens, Mistral Large 4 undercuts every American frontier model on price: GPT-6.1 Sol runs $2/$10, Sonnet 5.5 runs $2/$10, and Opus 5.5 runs $4/$20. Combined with EU-domiciled serving, the offer to European enterprises is frontier capability with GDPR-clean data residency at the lowest frontier price on the board. The open-weights promise within weeks would complete the package and put Mistral in direct competition with Kolibri and the open-weights leaders it benchmarked against for the European deployment dollar.

The Training Story: RL Without Saturation

The engineering details suggest Mistral treats this launch as a checkpoint, not a finish line. The 3,800-GPU pre-training run was followed by an RL effort running at roughly 3,000 GPUs producing about 33 billion tokens per day (around 16 billion trainable completions), and Mistral states the RL run is still ongoing with the model showing no saturation, meaning the public preview should improve in place. For a company that has been criticized for lagging the American labs, publishing a live-training frontier model with a promised open-weights release is a declaration that the gap is closing on Mistral’s terms.

What to Watch

Three things. First, the open-weights release by end of October: the license terms will determine whether Large 4 becomes the European self-hosting standard or a preview promise. Second, the cyber capability: an 82% cyber score with a lower refusal rate than its competitors is both the model’s differentiator and its regulatory liability, and how European regulators treat a model that is simultaneously “sovereign” and maximally capable of offensive security work is genuinely uncharted. Third, the Series D: a 3 billion euro raise on the back of a live-training preview means Mistral’s revenue needs to materialize in European enterprise deals, against Kolibri on sovereignty and American labs on price. The European AI market just got its first real competitive triangle.