Introducing Mistral Large 4 ("Le Chonk")

What it claims

Mistral Large 4 is a 1T-parameter, 52B-active, natively multimodal MoE. It is a hybrid instruct-and-reasoning model: one model covers instruction following, reasoning and agentic use. It is in public preview on Mistral Studio now, and weights are promised "by the end of the month." Mistral positions it as the strongest open-weight model built outside China, and as state-of-the-art among open models for cybersecurity, finance, law and visual grounding. Training data spans 160+ languages, including every official EU language.

How it works

The page gives no architecture detail beyond total and active parameter counts. Architecture, further benchmarks and post-training methodology are deferred to the weights release. What it does describe is the RL post-training system:

• Environments. A single run mixes task types through a shared, composable interface: chat, scientific problem solving, safety alignment, factuality and long-horizon tool use. The environments share code sandboxes, web search and external APIs.
• Verification. Each task combines whichever verifiers suit it: reward models, unit tests, LLM judges and static checks.
• Rollouts and training. An autoscaling actor fleet generates tens of thousands of rollouts in parallel, and training runs asynchronously alongside it. Rollout budgets reach millions of tokens across multiple context compactions. Mistral says it has "novel methods" to limit off-policy drift but does not describe them.
• Throughput. On about 3k GPUs, a run produces about 33B tokens/day, of which about 16B are trainable completion tokens after filtering and masking.
• Gains. Final gains are attributed to both SFT and RL. This is the same stack Mistral sells to customers as Mistral Forge.

flowchart LR
  E[Composable environments] --> A[Autoscaling actors]
  A --> R[Long rollouts]
  R --> V{Per-task verifier}
  V --> T[Async trainer]
  T -->|updated policy| A

Evidence

Cybersecurity

• On the Artificial Analysis vulnerability reproduce-and-patch test it scores 82%, the highest of any model. Claude Opus 5.5 and GPT-6 Astra score near zero on this test because they refuse it, so the lead reflects refusal behaviour as well as capability.
• It ranks top-5 overall on the AA Cyber Index.
• It solves 93% of Cybench's 40 challenges.

Coding

• DeepSWE v1.1: 61.7%.
• SWE-Atlas-QnA: 59.4%.
• Terminal-Bench 4: 28.3%.
• Coding Agent Index: 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
• In a blind Surge AI evaluation (1–5 scale), it scored 3.74. That puts it second of five: behind Claude Opus 5 (4.22) and ahead of GLM-5.3 (3.60), Kimi K3 (3.59) and GLM-5.2 (3.40).

Agents

• AutomationBench (657 business workflows): 59.9%, ahead of Kimi K3, MiMo-V2.6-Pro and DeepSeek V4 Pro.
• AA-Briefcase: 1,393 Elo, ahead of DeepSeek V4 Pro.

Vision

• Dense 200: 42% vs GPT-6 Astra's 41%, a one-point margin.

Knowledge work

• On vals.ai's legal and finance tasks it exceeds GPT-6 Astra. No numbers are given.
• On Harvey's Legal Agent benchmark it beats all open-source models.

Safety

• Lakera B3: resists 93.3% of attacks.
• KORA: 1.691 out of a maximum of 2.
• Its refusal rate on cyber prompts from JailbreakBench, StrongREJECT and AgentHarm is higher than all other OSS models.

Internal human evaluation vs GLM-5.3

• Preferred in CAD and STEM.
• On par or close in finance and coding.

How to use it

• Access: preview API on Mistral Studio today. Weights are promised by end of month, with self-deployment on private cloud or on-prem.
• Price: $1.36 per million input tokens and $4.18 per million output tokens.
• Hosting: available in multiple regions. This includes a European deployment operated end-to-end by Mistral, independent of other providers and under European law.
• Not stated: the page gives no licence, model ID, context length or hardware requirements for self-hosting.

Limits

• The model is still changing. This is a preview, the RL run is "still in flight," and Mistral expects "large and rapid improvements." Treat the numbers as a snapshot.
• The weights are not released and the licence is unstated. "Open-weight" is a promise as of this post.
• Cyber access differs by user. Red-teaming partners and state authorities get a version with reduced moderation and expanded cyber capabilities. The public preview is the more restricted version and refuses more on cyber prompts than other OSS models. Verify whether that blocks your security workflow.
• Several comparisons are thin:
• The STEM, CAD and finance human evaluations are internal and compare only against GLM-5.3.
• The vals.ai results come without figures.
• The vision lead over GPT-6 Astra is one point.
• The coding comparison is limited. Against closed models, coding is compared only to Claude Opus 5 in the human evaluation, where ML4 trails.