Mistral Large 4 Open Weights Released for 1-Trillion-Parameter Le Chonk Model

Mistral AI released the open weights for Mistral Large 4, nicknamed Le Chonk, on October 29, 2026. The model contains roughly one trillion total parameters and runs as a sparse Mixture-of-Experts architecture with 49 to 52 billion active parameters per token.

Mistral Large 4 Architecture and Release Details

The decoder-only transformer uses 72 layers, MLA attention, and a 1.6-billion-parameter Pixtral vision encoder that accepts text and image inputs while producing text output. It supports a native 1-million-token context window and more than 160 languages, including every official EU language. Training occurred from scratch on 3,800 to 4,000 NVIDIA Grace Blackwell GPUs inside Mistral’s European data centers.

Weights are distributed in FP8 and NVFP4 formats under the Mistral Ecosystem Open-Weight License, a permissive MIT derivative that allows broad fine-tuning and redistribution. Model-as-a-service hosting requires a separate commercial license. GGUF quantized versions and day-zero support for transformers, vLLM, and llama.cpp appeared on Hugging Face at the same time. The existing Mistral Studio API continues to operate for users who prefer managed inference.

Three reasoning modes—low, medium, and high—ship with high as the default. Post-preview reinforcement learning over three weeks incorporated customer feedback and lifted several benchmark scores before the open release.

Benchmark Results on Coding Agents and Multimodal Tasks

Mistral Large 4 leads or ties the top open-weight models on Terminal-Bench 4.0 at 41 percent, up from roughly 28 percent in the initial preview. It scores around 48 percent on the Artificial Analysis Coding Agent Index while consuming 47 percent fewer tokens than GLM-5.3. On cybersecurity evaluations the model reaches 79 to 82 percent on vulnerability reproduction and patching tasks and 93 percent on Cybench, with a high refusal rate on malicious prompts.

Multimodal performance shows competitive visual grounding on enterprise datasets in finance and law. The model ranks fifth on the Artificial Analysis Intelligence Index as of late October 2026 among all open systems. Continued reinforcement learning shows no signs of saturation, suggesting further gains are possible through additional training iterations.

Self-hosted deployment gives organizations direct control over data flows, eliminating vendor refusals common in closed models and supporting sovereign-AI requirements in regulated sectors. European training data and infrastructure further align with data-residency rules for government and enterprise users.

Practical Steps for Developers and Next Developments

Teams can download the checkpoints today and deploy them with existing inference engines to prototype coding agents or multimodal assistants without external API calls. Fine-tuning on domain data is permitted under the license, enabling specialized variants for internal tools or client projects. Organizations evaluating the model for production should test token efficiency and refusal behavior on their specific workloads before committing resources.

Mistral has indicated ongoing ecosystem work around inference optimizations and fine-tuning recipes. No dates have been set for specialized variants, yet the current release already positions the model as the strongest open-weight option outside China on coding and defensive cybersecurity benchmarks.