Mistral Large 4 Launches With 1T Parameters and 82% Cyber Score
Mistral Large 4 brings 1T parameters, a 1M-token context and 82% cyber score to Hugging Face, with open weights, vision and flexible local deployment options.
Summary
On October 29, 2026, Mistral released Mistral Large 4, or Le Chonk, on Hugging Face under the Mistral Ecosystem Open-Weight License, an MIT derivative restricting model-as-a-service offerings. The multimodal, multilingual Mixture-of-Experts model accepts text and images and generates text, with about 1 trillion total parameters, 49 billion active parameters and a 1 million-token context. Official FP8 and NVFP4 checkpoints join GGUF quants from Unsloth and llama.cpp, with day-one Transformers, vLLM, llama.cpp and TRL support.
Its 72-layer decoder has width 6,144 and 64 heads, combining compressed Multi-head Latent Attention with a DeepSeek Sparse Attention indexer selecting up to 2,048 positions. Layer one is dense; the remaining 71 route each token to four of 128 experts plus one shared expert. A 3:1 RoPE-to-NoPE pattern uses LeRope, while KV state is shared across layers. The 1.6 billion-parameter Pixtral vision encoder has 30 layers, width 2,048, 16 heads, 14 by 14 patches and spatial merge size 2.
As of release, ML4 ranked fifth among open-weight models on Artificial Analysis Intelligence. It scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4 and 49.8% on Coding Agent Index, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max; knowledge-work results were 59.9% on AutomationBench and 1,393 Elo on AA-Briefcase. Cyber results included 82% vulnerability reproduction and patching, top-five placement on Artificial Analysis Cyber Index, 93% across 40 Cybench challenges and 93.3% prompt-injection resistance on Lakera B3, plus internal utility for malware analysis, prioritization and detection rules. Reasoning offers low, medium or high but cannot be disabled. Deployment includes eight NVIDIA H200s for 131K context and 16 concurrent requests, or 2-bit GGUF on one 512GB Apple Silicon Ultra; vLLM’s NVFP4 path requires NVIDIA Blackwell GPUs.
Positives
- 82% vulnerability reproduction and patching was the highest reported result for that test within the Artificial Analysis Cyber Index.
- 93% of 40 Cybench challenges were solved, while Lakera B3 measured 93.3% resistance to prompt-injection attacks.
- 49.8% on Coding Agent Index put Mistral Large 4 ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.
- Day-one support covers Transformers, vLLM, llama.cpp and TRL, with FP8, NVFP4 and community GGUF variants available.
- Open weights let security teams control deployment and usage policies while directly measuring refusal behavior.
- 2-bit quants can run locally on one Apple Silicon Ultra system with 512GB of unified memory.
Risks & concerns
- The Mistral Ecosystem Open-Weight License limits model-as-a-service offerings despite being derived from MIT.
- Reasoning cannot be disabled, and selecting none still invokes the default high reasoning effort.
- vLLM’s NVFP4 checkpoint requires NVIDIA Blackwell GPUs, while FP8 fine-tuning was demonstrated on eight B300s.
- IQ2_XXS generation fell from 30 tokens per second at short contexts to 3 tokens per second at long contexts.
- The eight-H200 Inference Endpoints recipe limits context to 131K tokens, below the model’s advertised 1 million-token window.