# mistral-large-4-le-chonk-2026-10-06

## Veille

Blog post from **Mistral AI** (signed "By Mistral," 11-minute read), dated **October 6, 2026**, announcing the public preview of **Mistral Large 4** (ML4, nicknamed "the Chonk" by the publisher): a natively multimodal model with **1 trillion parameters, 49 billion of them active**, trained from scratch on **3,800 NVIDIA Grace Blackwell GPUs** in Mistral's European data centers. Preview API available on Mistral Studio; **open weights announced for the end of the month**.

## Titre Article

Introducing Mistral Large 4

## Date

2026-10-06

## URL

https://mistral.ai/news/mistral-large-4/

## Keywords

Mistral Large 4, ML4, the Chonk, Mistral AI, open weights, MoE model, multimodal, cybersecurity, AI sovereignty, Europe, agentic code, DeepSWE, Terminal-Bench, Cybench, Artificial Analysis Cyber Index, reinforcement learning, Mistral Forge, Harvey, vals.ai, prompt injection robustness, 160 languages, série D

## Authors

Mistral AI — éditeur de modèles d'IA ; billet de blog de l'entreprise (« By Mistral »).

## Ton

Profile: vendor product announcement aimed at technical decision-makers and enterprises, structured by capability domains (cybersecurity, agentic code, agentic workflows, multimodal, science and math, knowledge work, model safety), with benchmark tabs and four interactive demos. The text alternates between figures, named comparisons, and a sovereignty argument ("Forged in Europe. Built for AI sovereignty"). It uses a familiar nickname for the model and specifies the product's provisional status: preview, reinforcement learning "still ongoing," with architecture, additional benchmarks, and the post-training method promised alongside the weights release. The evaluations cited combine third parties (Artificial Analysis, vals.ai, Surge AI, Lakera, HarveyAI) and internal evaluations.

## Pense-betes

- **Model**: 1T parameters, 49B active, natively multimodal, hybrid instruction-and-reasoning model; more than **160 languages**, including all official EU languages. Trained with the training, customization, and reinforcement-learning environment offered to customers via **Mistral Forge**, with companies from finance, industry, logistics, pharma, and the public sector.
- **Timeline**: API preview today, weights before the end of the month. Before that, real-world testing (red teaming) with cybersecurity leads, verified partners, and state authorities, with reduced moderation and extended cyber capabilities.
- **Cybersecurity**: among the top five models on the AA Cyber Index and leading open models outside China; 82% on vulnerability reproduction plus fix, 93% on Cybench. The post explains that closed models' refusals block vulnerability research and incident response, and that losing access mid-crisis is a risk; it highlights private or on-premises deployment.
- **Code and agents**: 59.4% on SWE-Atlas-QnA, Coding Agent Index of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max; 59.9% on AutomationBench (657 workflows); 1,393 Elo on AA-Briefcase.
- **Multimodal**: visual and agentic capabilities combined (gigapixel satellite imagery, technical drawings); 42% versus 41% for GPT-6 Astra on Dense 200 (visual grounding).
- **Model safety**: 93.3% of attacks blocked on Lakera's B3; KORA score of 1.691 out of 2; refusal rate for malicious cyber requests, according to the post, higher than for all compared open models.
- **Reinforcement learning**: library of composable environments (code sandboxes, web search, APIs), combined verification (reward models, tests, LLM judges); with 3,000 GPUs, about **33 billion tokens per day**, of which ~16 billion are trainable; rollout budgets of several million tokens.
- **What's next**: ML4 is presented as the first step of the roadmap funded by the **série D de 3 Md€** round and as the base for a new generation of specialized models.
- ⚠️ **Scope**: preview; the detailed architecture and the post-training method are not published. The comparisons are Mistral's own, with competitor choices specific to the post, and several scores are given without the full table in the text (tabbed charts).

## RésuméDe400mots

Mistral AI announced on October 6, 2026 the public preview of Mistral Large 4, or ML4, nicknamed "the Chonk." The model has 1 trillion parameters, 49 billion of them active, is natively multimodal, and was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in the company's European data centers. The preview API is available on Mistral Studio; the weights will be released before the end of the month, following a real-world testing phase with cybersecurity leads, verified partners, and state authorities, who get access to a version with reduced moderation and extended cyber capabilities.

Mistral presents ML4 as close to the world's best open models and clearly ahead of any American or European open model, with state-of-the-art results among open models in cybersecurity, finance, and law, and results that surpass some closed models in visual grounding.

In cybersecurity, the model ranks in the top five of the Artificial Analysis Cyber Index and leads open models outside China. On the vulnerability reproduction-then-fix test, it scores 82%, the highest score cited, and 93% on Cybench. The post explains that Claude Opus 5.5 and GPT-6 Astra score close to zero on this test because they refuse the task, whereas defending software often starts with proving that a flaw is real. It concludes that this shows the value of open models, deployable privately or on-premises, under the organization's own policy.

In agentic code, ML4 reaches 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4, for a combined index of 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. A blind human evaluation by Surge AI ranks it second out of five (3.74), behind Claude Opus 5 (4.22). In agentic workflows, it scores 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. On knowledge work, vals.ai ranks it ahead of GPT-6 Astra in finance and law, and it outperforms other open models on Harvey's legal benchmark. The security section cites 93.3% of attacks blocked on Lakera's B3 benchmark and a higher refusal rate for malicious cyber requests than the open models it is compared against.

On methodology, Mistral describes a composable reinforcement-learning library in which sandboxes, web search, APIs, and verifiers (reward models, tests, judges) combine. With 3,000 GPUs, a training run produces about 33 billion tokens per day. ML4 is the first step of the roadmap funded by the série D de 3 Md€ round and will serve as the base for specialized models. Pricing is $1.36 per million input tokens and $4.18 per million output tokens.

## GrapheDeConnaissance

- Mistral AI —publie→ Mistral Large 4 (TECHNOLOGIE, 0.98)
- Mistral Large 4 —mesure→ 1 000 milliards de paramètres dont 49 milliards actifs (MESURE, 0.97)
- Mistral Large 4 —utilise→ GPU NVIDIA Grace Blackwell (TECHNOLOGIE, 0.95)
- Mistral AI —recommande→ déployer ML4 en privé ou sur site pour les opérations de sécurité (AFFIRMATION, 0.85)
- Mistral Large 4 —mesure→ 82 % sur reproduction puis correction de vulnérabilité (AA Cyber Index) (MESURE, 0.95)
- Mistral Large 4 —mesure→ 93 % sur Cybench (MESURE, 0.95)
- Mistral Large 4 —surpasse→ Claude Opus 5.5 (TECHNOLOGIE, 0.85)
- Mistral Large 4 —surpasse→ GPT-6 Astra (TECHNOLOGIE, 0.85)
- Mistral Large 4 —surpasse→ DeepSeek-V4-Pro (TECHNOLOGIE, 0.88)
- Claude Opus 5 —surpasse→ Mistral Large 4 (TECHNOLOGIE, 0.9)
- Mistral Large 4 —mesure→ 61,7 % sur DeepSWE v1.1 (MESURE, 0.95)
- Mistral Large 4 —mesure→ 3,74 sur 5 à l'évaluation humaine Surge AI (MESURE, 0.93)
- Mistral Large 4 —mesure→ 93,3 % d'attaques repoussées sur Lakera B3 (MESURE, 0.93)
- Mistral AI —affirme_que→ les refus des modèles fermés bloquent la recherche de vulnérabilités et la réponse à incident (AFFIRMATION, 0.9)
- Mistral Large 4 —utilise→ Mistral Forge (TECHNOLOGIE, 0.9)
- Mistral Large 4 —utilise→ apprentissage par renforcement (CONCEPT, 0.95)
- série D de 3 Md€ —permet→ la feuille de route de développement du modèle (AFFIRMATION, 0.88)
- Mistral Large 4 —permet→ souveraineté de l'IA (CONCEPT, 0.85)

---
Canonical: https://www.thekb.eu/en/fiches/mistral-large-4-le-chonk-2026-10-06/