Mistral Large 4: Trillion-Parameter Open-Weight Multimodal MoE

Mistral AI's largest model to date (nicknamed Le Chonk), announced as an open-weight release in the official blog on 10 29: a native multimodal MoE with approximately 1 trillion total parameters, activating 490 billion per token (520 billion including embedding and output layers), with 4 of 128 routed experts activated per pass, a context window of 1000,000 tokens, and support for text and image inputs. Instruction, reasoning, coding, and agentic tasks share a single set of weights; reasoning intensity can be switched per request among low / medium / high levels, with support for function calling and strict JSON Schema output. Officially, it scores 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, while using 47% fewer tokens per task on average; it scores 26 on Intelligence Index v5, the highest among Western open-weight models, but lower than leading Chinese open-source models. The checkpoint includes a built-in DSpark speculative decoding draft head, with official claims of up to 1.9x throughput improvement.

Original title: mistralai/Mistral-Large-4-1T-A52B

Mistral AI's largest model to date (nicknamed Le Chonk), announced as an open-weight release in the official blog on 10 29: a native multimodal MoE with approximately 1 trillion total parameters, activating 490 billion per token (520 billion including embedding and output layers), with 4 of 128 routed experts activated per pass, a context window of 1000,000 tokens, and support for text and image inputs. Instruction, reasoning, coding, and agentic tasks share a single set of weights; reasoning intensity can be switched per request among low / medium / high levels, with support for function calling and strict JSON Schema output. Officially, it scores 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, while using 47% fewer tokens per task on average; it scores 26 on Intelligence Index v5, the highest among Western open-weight models, but lower than leading Chinese open-source models. The checkpoint includes a built-in DSpark speculative decoding draft head, with official claims of up to 1.9x throughput improvement.

Why it matters A rare Western trillion-parameter open-weight model, officially claiming coding agent performance on par with leading open-source models with lower token usage

Upvotes 78 · Likes 1646 · Downloads 12541

Reasons

  • A rare trillion-parameter general-purpose model among open weights, covering vision, reasoning, coding agents, and tool calling with a single set of weights
  • Officially claimed to excel in cybersecurity, finance, and legal tasks, as well as visual grounding, supporting 160 languages
  • Day-one support from vLLM, SGLang, transformers, and llama.cpp; fine-tuning possible with Axolotl, Unsloth, and TRL; an NVFP4 quantized version for Blackwell is also available

Usage

  • vLLM: Use the official image vllm/vllm-openai:mistral-large-4, with 8-card tensor parallelism, adding --tool-call-parser mistral --reasoning-parser mistral; add --speculative-config to enable DSpark speculative decoding
  • Officially recommended temperature 1.0, top_p 0.95; use reasoning_effort=high for complex tasks and agents, medium for daily use, and low for quick chats
  • If not self-hosting, call mistral-large-4 via the Mistral API, or use it with the official open-source coding assistant Mistral Vibe

Risks

  • Licensed under the Mistral Ecosystem Open-Weight License (a modified MIT), allowing commercial use, but operating a Model-as-a-Service business requires a separate commercial license
  • Example deployments require a 8-card node, making it impossible to run on personal devices
  • Performance metrics come from Artificial Analysis indices and the official blog cited by the vendor
  • This entry has not been tested; refer to the model card for actual performance

Open

Hugging Face Trending

Explore the latest AI Radar updates