What Is Mistral Large 4? 'Le Chonk' Open-Weight Model
The short answer
Mistral Large 4 is Mistral AI’s new flagship: a 1.05-trillion-parameter, open-weight-bound multimodal model released as a public preview on October 6, 2026. It activates 52 billion parameters per token, reads images through a 1.6B vision encoder, handles a 1 million-token context, and costs $0.68 / $2.09 per million input/output tokens at its preview sale price. Mistral calls it “ML4”, officially “le Chonk”, and says the weights will be released by the end of October 2026.
Key facts
| Mistral Large 4 | |
|---|---|
| Released | October 6, 2026 (public preview, API model mistral-large-4) |
| Architecture | Granular mixture-of-experts: 1.05T total, 52B active, 1.6B vision encoder |
| Context | 1M tokens |
| Input | Text and images |
| Price (preview sale) | $0.68 input / $0.07 cached / $2.09 output per MTok |
| List price | $1.36 / $0.14 / $4.18 per MTok |
| Weights | Promised by end of October 2026 |
| Training | From scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centres |
| Languages | Training data in 160+ languages, all official EU languages |
What Mistral says it is good at
Cybersecurity. Mistral’s headline claim: a top-five rank on the Artificial Analysis Cyber Index, 82% on its reproduce-and-patch-a-real-vulnerability test (the highest of any model, according to Mistral) and 93% on Cybench’s 40 capture-the-flag challenges. Mistral argues that closed models such as Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduction test because they refuse it — a feature for most users and a limitation for defenders. Vetted partners currently get a version with reduced moderation and expanded cyber capabilities.
Agentic coding. 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4 and a 49.8% Coding Agent Index score. For comparison, Claude Sonnet 5.5 is reported at 70.6% on Terminal-Bench 4.0, so ML4 is not a frontier closed-model replacement for terminal agents.
Business workflows. 59.9% on AutomationBench (657 workflows across Gmail, Sheets, Slack and Salesforce) and 1,393 Elo on AA-Briefcase, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro by Mistral’s count.
Vision. Mistral reports 42% on Dense 200 visual grounding, one point above GPT-6 Astra’s 41%.
How it compares on price
At the sale price, ML4 costs about a third of Claude Sonnet 5.5 or GPT-6.1 Sol ($2 / $10) on input and a fifth on output. Even at the $1.36 / $4.18 list price it is cheaper than every US closed model above the small tier. Its competition is other open-weight frontier models: Xiaomi’s MiMo-V2.6-Pro at $0.435 / $0.87, and Kimi K3 or GLM 5.3 on third-party hosts. See our current API prices.
Who should care
- European companies with sovereignty requirements: Mistral runs a European deployment end-to-end under EU law, and the open weights will allow fully private hosting.
- Security teams: the cyber results and the absence of provider-level refusals once self-hosted are the main reason to evaluate it.
- Cost-sensitive agent builders: a 1M context at under $1 input is attractive for document-heavy agents.
Caveats. All benchmark figures are Mistral’s own as of October 10, 2026. The licence for the open weights has not been detailed yet, and a 1.05T-parameter model needs a multi-GPU server to self-host. Mistral says ML4 will be the base for a new generation of specialised Mistral models.
Related: best open-weight AI models ranked by value and Claude Haiku 5.5 vs GPT-6 Luna vs Gemini 3.8 Flash for the other frontier-lab launch of October 7, 2026.
Last verified: October 10, 2026.