Mistral releases weights for Mistral Large 4 (Le Chonk), focusing on coding agents and cybersecurity

On 10 29, Mistral released the weights for its flagship model, Mistral Large 4 (nicknamed Le Chonk). It is a Mixture-of-Experts model with approximately 1 trillion parameters, supporting text and image inputs and text outputs, with a context window of 100 thousand tokens; it launched as a preview version three weeks ago, after which reinforcement learning continued, and officials claim significant improvements in coding and cybersecurity capabilities. The license is a modified MIT: free use, modification, fine-tuning, and commercial use are permitted, but offering it externally as "Model-as-a-Service" requires a separate commercial license.

On 10 29, Mistral released the weights for its flagship model, Mistral Large 4 (nicknamed Le Chonk). It is a Mixture-of-Experts model with approximately 1 trillion parameters, supporting text and image inputs and text outputs, with a context window of 100 thousand tokens; it launched as a preview version three weeks ago, after which reinforcement learning continued, and officials claim significant improvements in coding and cybersecurity capabilities. The license is a modified MIT: free use, modification, fine-tuning, and commercial use are permitted, but offering it externally as "Model-as-a-Service" requires a separate commercial license.

Confidence: High (100) · Event time 2026-10-29

Reasoning

  • A rare trillion-parameter-scale model among open-weight models, trained entirely in Europe, allowing enterprises and governments to run it in their own data centers without relying on APIs. Official claims: On Cognition's private evaluation FrontierCode 2.0, it scored 19.7%, ranking first among open-weight models; on Terminal-Bench 4.0, it improved from 28% in the preview version to 41%; on the Artificial Analysis Coding Agent Index 48%, it is roughly on par with GLM-5.3 but uses 47% fewer tokens per task.
  • Cybersecurity is a key selling point (official claim): On the Artificial Analysis Cybersecurity Index 51%; on malicious web requests in JailbreakBench, StrongREJECT, and AgentHarm, the average rejection rate is 96.6%, the highest among open models; additionally, a variant specialized for penetration testing and red teaming is being previewed for vetted partners such as Orange Cyberdefense. Mistral also acknowledges it may aid real-world attacks and will continue adjusting protections over the coming months.
  • Third-party results are more sober: According to Artificial Analysis as reported by The Decoder, it scored 26 in the new Intelligence Index v5 (preview version 25), lower than DeepSeek V4.1 Flash and GLM-5.3 which scored 32; on the Cybersecurity Index, DeepSeek V4.1 Flash (57%) and MiMo-V2.6-Pro (56%) also rank ahead of it. Its advantages are speed and price: it outputs 161 tokens per second, averaging 2.06 per index task, cheaper than GLM-5.3 but significantly more expensive than DeepSeek V4.1 Flash.

Usage

  • Weights are available in the Mistral Large 4 collection on Hugging Face (mistralai/Mistral-Large-4-1T-A52B, with an NVFP4 version also available); day-one support includes transformers, vLLM, SGLang, and llama.cpp (GGUF quantization provided by ggml-org); reasoning intensity reasoning_effort can be set to low / medium / high, defaulting to high, and reasoning cannot be disabled.
  • High hardware requirements: According to SGLang documentation cited by The Decoder, the standard version requires 8 B300 GPUs; the Hugging Face blog states that 2-bit quantization can run on a single Apple Silicon Ultra with 512GB of memory, and provides a deployment solution using 8 H200 GPUs on Inference Endpoints.
  • For those who do not want to self-host, options include Mistral Studio, Mistral API, OpenRouter, Microsoft Foundry, or NVIDIA hosted endpoints; official pricing is 0.70 per million input tokens, 0.07 for cached input, and 2.10 for output.

Risks

  • Inconsistent claims regarding active parameters: The Mistral repository name, The Decoder, and SiliconANGLE state 520 billion, while the Hugging Face blog states approximately 490 billion.
  • Discrepancies in several numerical metrics: The 82% cited in the Hugging Face blog refers to the "reproduce and patch vulnerabilities" item in the Cybersecurity Index, with a total score of 51%; the B3 attack resistance score of 93.3% cited is the preview version result mentioned by Mistral (official claim for the final version is 97.1%); Terminal-Bench 4 reports 28.3%, a significant difference from Mistral's 41%, with test settings unspecified.
  • Most coding and security scores are from evaluations selected by Mistral; endorsements from partners such as Booz Allen and Cisco are considered testimonials; the full technical report is officially stated to be "coming soon".
  • The Decoder points out that Mistral's comparisons did not include Kimi K3.1 Pro (weights not yet released, Moonshot plans to release before 11 4), which scored 67% on Terminal-Bench 4.0.

Open · Mistral AI News · The Decoder · TechCrunch · SiliconANGLE

Hugging Face

Explore the latest AI Radar updates