Mistral Large 4 Open Weights: License, Hardware and Price
The short answer
Mistral Large 4 is now a downloadable open-weight model. On October 29, 2026 Mistral released the weights of an improved Mistral Large 4 — the 1-trillion-parameter, natively multimodal model it previewed on October 6 — under a modified MIT license, and set the API price at $0.70 input / $0.07 cached / $2.10 output per million tokens. You need an eight-GPU Blackwell server (or two H200 nodes) to run it yourself. For what the model is, see our Mistral Large 4 explainer.
What shipped
| Mistral Large 4 (open weights, Oct 29, 2026) | |
|---|---|
| Checkpoints | mistralai/Mistral-Large-4-1T-A52B (FP8, ~1.06 TB) and -NVFP4 (~610 GB) on Hugging Face |
| Architecture | ~1T MoE; 49B active per token (52B with embeddings/output); 128 routed experts (4 active) + 1 shared; 72 layers |
| Context | 1,048,576 tokens (vLLM validated up to 256K) |
| Input / output | Text and images in, text out |
| Reasoning | reasoning_effort low / medium / high (default high) |
| Extras | Embedded DSpark speculative-decoding draft head (3-token blocks) |
| License | Mistral Ecosystem Open-Weight License (modified MIT) |
| API price | $0.70 / $0.07 cached / $2.10 per MTok |
| Also on | Mistral Studio, OpenRouter, Microsoft Foundry, NVIDIA endpoints and NIM |
What changed since the preview
Mistral says three more weeks of reinforcement learning produced the released checkpoint. Vendor-reported results:
- Terminal-Bench 4.0: 41%, up 13 points from the preview’s 28%, with 30% fewer output tokens — the top open-weight score Mistral reports.
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task.
- Cognition FrontierCode 2.0: 19.7%, which Mistral says leads.
- Artificial Analysis Cyber Index: 51%, and 97.1% attack resistance on the B3 Agent Security Benchmark (93.3% in preview).
- BrowseComp 84.5% and Finch (FinWorkBench) 73.3%, second among open models to Kimi K3.
Closed frontier models remain ahead on terminal agents: Claude Sonnet 5.5 reports 70.6% on Terminal-Bench 4.0. Treat these as vendor numbers until independent runs land.
The license, in plain terms
The Mistral Ecosystem Open-Weight License is MIT with one carve-out. You may use, modify, fine-tune and commercialise the model. If your business is selling access to the model as a service — an inference API or hosted endpoint — you need a separate commercial license from Mistral first. Running it inside your own product, internal tools or agents does not trigger that condition. Compare that with Meta’s Muse Spark Open, which requires attribution and an agreement for model-as-a-service companies above $20M revenue.
What it takes to self-host
vLLM’s official recipe validates:
- NVFP4 on 8x B200 or 8x B300, or a 4-GPU GB300 tray — about 76 GiB of weights per GPU on eight GPUs.
- FP8 on 2x 8 H200 (tensor parallel 16 or pipeline 2 × tensor 8) — about 128 GiB of weights per GPU on eight GPUs.
- AMD Instinct MI350X, eight GPUs, with the ROCm image and an MXFP4 checkpoint (~578 GB).
You need the dedicated vllm/vllm-openai:mistral-large-4 or lmsysorg/sglang:mistral-large-4 container images; as of November 1, 2026 no regular vLLM release or nightly includes support. On 8x GB300 with NVFP4 and DSpark, vLLM measured about 110 output tokens per second per GPU at 8K context while keeping 100 tokens per second per user.
At $0.70/$2.10 on the API, self-hosting pays only at sustained high volume or when data cannot leave your infrastructure. For a provider-by-provider view, see the cheapest inference providers for open-weight models and our current API prices.
Last verified: November 1, 2026.