The Lab Notes

Mistral

Mistral opens a Large 4 preview and promises open weights in October

Mistral launched a public preview of Mistral Large 4, a 1-trillion-parameter model, and says downloadable weights follow by the end of the month.

The Lab Notes ·

A purple block-shaped server module with a pink top, an orange perforated panel on its left end and a brass heat sink with fine fins pointing right, set on a black background with thin white arcs on the left and a repeating pattern of pink, purple and yellow server racks on the right.

Mistral opened a public preview of Mistral Large 4, a model with 1 trillion parameters, and says it will release the open weights by the end of October. Open weights are the trained model files, released for download.

Mistral announced the preview Oct. 6, 2026, in its Introducing Mistral Large 4 post. This is a preview, not general availability. Mistral’s docs changelog entry of the same date lists the preview API on Mistral Studio under the model ID mistral-large-4 and says “Open weights are coming soon.” The weights are not out yet.

Mistral Large 4 has 1 trillion parameters and a 1 million token window

Mistral calls Mistral Large 4, nicknamed “le Chonk,” its largest and most capable model. It is natively multimodal. The changelog says it is a mixture-of-experts design, in which only part of the network runs for each token. Here that part is 52 billion active parameters. The context window is 1 million tokens. Launch pricing is 50% off for two weeks, per the changelog.

Mistral trained the model on 3,800 Nvidia GPUs in Europe

Mistral said it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters. The training data covered more than 160 languages, including every official EU language. Mistral said it plans a European deployment run end to end by the company under European law.

Mistral says the model leads open-weight models on cyber tests

Mistral reports that the model ranks among the top five globally on the Artificial Analysis Cyber Index and leads open-weight models developed outside China.

On one test in that index, which asks a model to reproduce a real vulnerability in open-source software and patch it, Mistral said the model scores 82%, the highest of any model. It said Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task. It also solves 93% of Cybench’s 40 challenges, Mistral said.

Until the weights ship, Mistral said it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities. They get the same model with reduced moderation and expanded cyber capabilities.

Mistral reports its own coding, agent and safety figures

The preview scores 61.7% on DeepSWE v1.1, according to Mistral. Every figure below is Mistral’s report of results at launch on Oct. 6.

  • Coding: 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. Mistral put its combined Coding Agent Index at 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
  • Human rating: in a blind Surge AI coding evaluation on a 1-to-5 scale, the preview ranked second of five at 3.74. Claude Opus 5 scored 4.22 and GLM-5.3 3.60.
  • Agents: 59.9% on AutomationBench, which covers 657 business workflows, ahead of Kimi K3 and DeepSeek V4 Pro, Mistral said.
  • Safety: it resists 93.3% of attacks on Lakera’s public B3 benchmark, which Mistral said is the highest among the competitors it compared. Mistral said its cyber-prompt refusal rate is higher than all open-source models it compared.

Analysis

We think two things decide how this launch is judged. The first is the weights date. Downloadable files let outside researchers check the numbers above. Mistral has promised them by the end of October.

The second is the cyber refusal comparison. Security work that closed rivals decline is a selling point for an open-weight model. We would watch whether outside testers reproduce Mistral’s refusal comparison once they can run the model.