What Was Released
Mistral published open weights for Mistral Large 4 on October 29, 2026, in a post titled "Le Chonk: Built for Coders, Trusted by Defenders". The API preview had launched on October 6. The Hugging Face repo is mistralai/Mistral-Large-4-1T-A52B, with a separate NVFP4 checkpoint for Nvidia Blackwell GPUs and GGUF quantizations. The model card describes about 1 trillion total parameters and 49B active per token (52B counting embedding and output layers), 128 routed experts with 4 active plus 1 shared expert, a vision encoder, and a 1M-token context.
Step 1: Check the License First
The weights use the Mistral Ecosystem Open-Weight License, a modified MIT license. Commercial use is allowed, but running a model-as-a-service business on it needs a separate commercial licence from Mistral. Internal tools and products that use the model are fine. Reselling API access to it is not.
Step 2: Size the Memory
Memory follows total parameters, not active ones, so plan for all 1T weights. See active parameters for why.
| Precision | Weights alone, approximate | Typical hardware |
|---|---|---|
| FP8 (as published) | About 1 TB | 8 data-center GPUs; Mistral's vLLM and SGLang examples use tensor parallel across 8 |
| 4-bit (NVFP4 or GGUF Q4) | About 500 GB (press estimate) | 4 Blackwell GPUs, or two 512 GB Mac Studios |
| Lower-bit GGUF | Less, with quality loss | A single 512 GB Mac Studio, with little room for long context |
Mistral doesn't state a VRAM figure. These numbers are arithmetic from the parameter count, and you need extra memory for the context cache. At long contexts that cache gets large. Meta's 816B Muse Spark Open, which is similar in size, needed two M5 Ultra Mac Studios for Apple's 4-bit tests. See Muse Spark Open hardware requirements.
Step 3: Pick the Checkpoint
- FP8 main repo: for 8-GPU servers with vLLM or SGLang.
- NVFP4: for B200, B300 or other Blackwell GPUs. It roughly halves memory against FP8.
- GGUF: for llama.cpp, LM Studio and Ollama on Macs or large CPU-RAM servers.
Step 4: Serve It
On a GPU server, follow the model card's vLLM or SGLang example with tensor parallel set to 8. On a Mac, load a GGUF in llama.cpp or LM Studio. For more than one Mac, use a distributed runtime such as MLX's. Start with a short context, such as 8K to 32K tokens, and raise it only once speed is acceptable. The 1M window is the model's limit, not something a single desktop will serve quickly. For tools, see Ollama vs LM Studio vs LocalAI.
Step 5: Decide If Local Is Worth It
Mistral's API charges $0.70 per million input tokens and $2.10 per million output, with cached input at $0.07. It is also on Mistral Studio, OpenRouter, NVIDIA NIM and Microsoft Foundry. At those prices, a $20,000-plus local setup pays off only with heavy, steady use, or when data can't leave your network. For the comparison with other open models, see Mistral Large 4 vs Muse Spark Open vs Reflection Beam.
What Mistral Reports
Mistral says Large 4 scores 26 on the Artificial Analysis Intelligence Index v5, which it calls the highest for a Western open-weight model, and Artificial Analysis lists that score. Its other figures are self-reported: 48% on the AA Coding Agent Index, 41% on Terminal-Bench 4.0 (up from 28% for the preview), 51% on the AA Cyber Index and 84.5% on BrowseComp.