French AI Firm Mistral Releases Open-Weight Model Mistral Large 4: Full Weights, Architecture Details and Benchmarks Available
Mistral Large 4 officially launched on Hugging Face, revealing its 1T-parameter MoE architecture and cybersecurity benchmarks, enabling local deployment.
French artificial intelligence company Mistral officially released its latest multimodal large language model, Mistral Large 4 (codenamed "Le Chonk"), on the Hugging Face platform on October 29, 2026. The model features approximately 1 trillion total parameters and 49 billion active parameters, utilizing a Mixture-of-Experts (MoE) architecture that accepts text and image inputs. This release marks the transition from preview to fully open weights, allowing developers to access official checkpoints in FP8 and NVFP4 formats.
Technically, Mistral Large 4 combines sparse attention mechanisms with a Pixtral vision encoder. Its text decoder comprises 72 layers, employing Multi-head Latent Attention (MLA) and a DeepSeek Sparse Attention (DSA) indexer, which selects up to 2,048 key positions for each query. This design aims to balance computational efficiency and performance through two forms of sparsity: selecting feed-forward network experts and selecting context positions. Additionally, the model uses YaRN-scaled RoPE (LeRope) to optimize long-context processing.
Mistral highlights the model's defensive capabilities in cybersecurity. According to Artificial Analysis's Cyber Index test, Mistral Large 4 scored 82% in reproducing and patching vulnerabilities in open-source software, ranking among the top five globally. Mistral notes that many closed frontier models refuse such tasks due to safety filters, whereas open weights allow users to control deployment policies, which is critical for security teams needing to reproduce vulnerabilities for remediation. The model also shows competitive performance in coding agents and multimodal understanding benchmarks, achieving 61.7% on DeepSWE v1.1.
The model is distributed under the Mistral Ecosystem Open-Weight License (an MIT derivative with restrictions on model-as-a-service offerings). Day-0 support is available for Transformers, vLLM, and llama.cpp, with community-provided GGUF quantizations for local execution. This release offers industrial enterprises seeking to avoid vendor lock-in and deploy high-performance code and security analysis models in private environments a new verifiable option.