Mistral Large 4 Released on Hugging Face with 1M Context and Top Cybersecurity Scores

• Mistral Large 4 (Le Chonk) is a multimodal Mixture-of-Experts model with 1T total parameters and 49B active parameters, released on Hugging Face on October 29, 2026.
• The model achieves an 82% score on the Artificial Analysis Cyber Index, the highest in its category, and ranks 5th on the Artificial Analysis Intelligence index among open-weight models.
• Architecture includes 72 layers, Multi-head Latent Attention (MLA), and DeepSeek Sparse Attention (DSA), supporting a 1M token context window.
• Day-0 inference support is available via transformers, vLLM, and llama.cpp, with official FP8 and NVFP4 variants and GGUF quantizations provided by Unsloth and the llama.cpp team.

Performance and Benchmarks

• Coding Agent Index score of 49.8%, outperforming DeepSeek V4 Pro 0813 and Qwen3.8 Max.
• Resolves 93% of 40 security contest problems on Cybench and demonstrates 93.3% resistance to prompt injection on the Lakera B3 benchmark.
• DeepSWE v1.1 score of 61.7% and SWE-Atlas-QnA score of 59.4%.

Deployment and Hardware

• Inference Endpoints on 8 x NVIDIA H200 GPUs achieve ~65 tok/s generation for single requests and ~480 tok/s total generation under load with 1,000 concurrent conversations.
• Local execution on a single M3 Ultra with Q2_K quantization yields ~90 tok/s prefill and 15 tok/s generation.
• vLLM deployment supports NVFP4 checkpoints on NVIDIA Blackwell GPUs and utilizes DSpark drafter for speed improvements.