Mistral's largest open-weight model to date: a native multimodal MoE with approximately 1T total parameters and about 52B active parameters per token (or 49B excluding embedding layers). It accepts text and images as input and outputs text, supporting a 100-token context window. Instruction following, reasoning, coding, and agentic tasks are unified within a single set of weights, with reasoning intensity switchable per request across three levels: low, medium, and high.
Upvotes 300 · Likes 1221 · Downloads 278
Reasons
- Officially, the Artificial Analysis Coding Agent Index 48% is roughly on par with the leading open-source GLM-5.3, but uses 47% fewer tokens per task on average; the Artificial Analysis Intelligence Index v5 scores 26, the highest among Western open-source models, though officials acknowledge that leading Chinese open-source models have higher overall scores.
- With 128 routing experts, 4 are activated per token along with 1 shared experts, and the model was trained from scratch on 3,800 Grace Blackwell GPUs; officials claim outstanding performance in cybersecurity, finance, law, and visual grounding, with support for native function calling and strict JSON Schema output, covering 160 languages.
- At launch, it included vLLM, SGLang images, Transformers, llama.cpp GGUF, as well as fine-tuning guides for Axolotl, Unsloth, and TRL, plus an NVFP4 version optimized for Blackwell; it was released on 29 10 alongside the Hugging Face blog post, having not been included in previous editions. The rising score of 300 reflects only attention levels.
Usage
- vLLM (officially recommended):
docker pull vllm/vllm-openai:mistral-large-4, launch with--tensor-parallel-size 8 --tool-call-parser mistral --reasoning-parser mistral --enable-auto-tool-choice; add--speculative-config '{"method": "dspark", "num\_speculative\_tokens": 3}'to enable built-in DSpark speculative decoding, which officials claim improves throughput by up to 1.9x. - SGLang: Use the
lmsysorg/sglang:mistral-large-4image, runsglang serve --tp-size 8 --speculative-algorithm DSPARK; use unified sampling with temperature 1.0 and top_p 0.95. Setreasoning\_effort="high"(default) for complex and agentic tasks,mediumfor everyday use, andlowfor quick responses. - You can also directly call the
mistral-large-4model via the Mistral API, or connect a local vLLM instance within the official coding agent, Mistral Vibe.
Risks
- Official examples deploy with tensor parallelism across 8 GPUs, making it unrunnable on personal devices; the NVFP4 version is exclusively for Blackwell GPUs.
- The license is the Mistral Ecosystem Open-Weight License (a modified MIT license): commercial use is permitted, but Model-as-a-Service (MaaS) businesses require a separate commercial license.
- The model card provides only two composite indices; other comparisons are found in images and blog posts, all citing official leaderboard results.