One-sentence summary
Mistral Large 4 (Le Chonk) opened weights at about 1.05T total and ~52B active parameters, leading western open weights on coding and defensive cybersecurity, while the company concedes Chinese open-weight models score higher on the aggregate intelligence index.
Background
On October 29, 2026, France's Mistral AI released the weights of its flagship Mistral Large 4 (ML4, codename Le Chonk). First previewed on October 6, the model continued to scale reinforcement learning on preview feedback. Mistral positions it as one of the strongest western open-weight models, with emphasis on coding agents and defensive cybersecurity.
Key Point 1: Architecture and Training
- Granular MoE: 128 routed experts, 4 active per token, plus 1 shared expert.
- About 1.05T total parameters and ~52B active per token (49B excluding embeddings and output).
- Native multimodal: text + image input, text output, with a 1.6B Pixtral-style vision encoder.
- Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters; natively fluent in 160+ languages.
Key Point 2: Benchmarks and Coding Agents
- Cognition's private FrontierCode 2.0 score 19.7%, leading open weights ahead of GLM-5.3, Kimi K3, DeepSeek V4.1 Flash.
- Artificial Analysis Coding Agent Index 48%, roughly tied with GLM-5.3 but using 47% fewer tokens per task.
- Terminal-Bench 4.0 score 41% (+13 points from preview), leading open weights with 30% fewer output tokens.
- Surge Code blind pairwise: second only to Claude Opus 5.5, ahead of GLM-5.3, Kimi K3, DeepSeek V4.1 Flash.
Key Point 3: Cybersecurity and License
- Artificial Analysis Cyber Index 51%; average refusal rate 96.6% on malicious cyber prompts across JailbreakBench, StrongREJECT, AgentHarm.
- Released under the Mistral Ecosystem Open-Weight License (modified MIT): free to use, modify, fine-tune and commercialize, but MaaS operators need a separate commercial license.
- API pricing: input $0.70/M, cached input $0.07/M, output $2.10/M; an NVFP4 quantized checkpoint targets Blackwell GPUs.
Key Point 4: What It Means for Open Weights
Mistral scores 26 on Intelligence Index v5, the highest among western open weights, but explicitly notes Chinese open-weight models score higher on the aggregate index. Le Chonk's value is not total domination; it gives western enterprises a self-hostable, commercially licensable large-weight model strong enough for coding and cybersecurity.
Sources
- Unite.AI: Mistral's 1-Trillion-Parameter Le Chonk Checkpoint Lands on Hugging Face
- IT Home (via 163.com): Mistral AI open-sources Large 4 weights