Mistral Large 4
Mistral Large 4 accepts text and images and returns text. Use mistral-large-4 on the hosted API or the released mistralai/Mistral-Large-4-1T-A52B checkpoint; deployment and licensing conditions differ.
Why it matters
Choose hosted API or self-managed inference, then evaluate hardware, licensing, total task cost and verified output quality.
API preview and architecture
The hosted model page retains its Public Preview label and 1M context. The current checkpoint card describes about 1T total parameters, 49B active per token and 52B including embeddings and output layers, plus a 1.6B vision encoder. The card explicitly explains the previously differing active counts.
Introductory and base token rates
Checked November 4, 2026: Mistral’s current USD model page lists $0.70 input, $0.07 cached input and $2.10 output per million tokens. These current displayed rates replace the earlier two-week launch-discount description; verify currency and current provider billing before sending requests.
Weights and license remain a separate gate
The released FP8 checkpoint is mistralai/Mistral-Large-4-1T-A52B, with an official NVFP4 sibling. The October 29 Hugging Face article documents both. Mistral’s model card identifies the Mistral Ecosystem Open-Weight License, a modified MIT license permitting commercial and non-commercial use while requiring a separate commercial license for a Model-as-a-Service business. Review the actual license and your deployment terms.
How to assess the preview
Start with representative tasks and matched tool, effort and latency budgets. Mistral links an official CTF speedrun demonstration, which is not a general benchmark or proof of safe production behavior. Validate the API contract and rollout terms before replacing an existing model.
Current specification and weight-release boundary
At the October 21 checkpoint, the sources described upcoming weights. That historical state is superseded by the released model card and October 29 FP8/NVFP4 release. Hosted access, downloadable artifacts and hardware/runtime remain separate checks; GetLLMs has not run this model.
- Current official source
Read the exact release or provider scope.
Useful first jobs
- Trial a multimodal API for document and coding tasks.
- Evaluate structured outputs and tool use with a defined harness.
- Evaluate self-managed inference using the released checkpoints and their license and hardware requirements.
Frequently asked questions
Can I download Mistral Large 4 weights today?
Yes. Official FP8 and NVFP4 checkpoints are released on Hugging Face. Start with mistralai/Mistral-Large-4-1T-A52B and review the Mistral Ecosystem Open-Weight License, including the separate Model-as-a-Service licensing requirement.
Is the launch price permanent?
No. The earlier two-week launch offer is historical. Checked November 4, 2026, the current USD model page lists $0.70 input, $0.07 cached input and $2.10 output per million tokens; check live billing before budgeting.
Sources and evidence
Checked November 4, 2026 against current Mistral documentation and its Hugging Face model card; the October 29 Hugging Face release article corroborates released FP8/NVFP4 variants. Earlier upcoming notices are historical.
- Mistral Large 4 model card ↗
Mistral · Primary source
- Mistral changelog ↗
Mistral · Primary source
- Mistral Large 4 announcement ↗
Mistral · Primary source
- Current model display ↗
Mistral · Primary source
- Research display ↗
Mistral · Primary source
- Released model card ↗
Mistral · Primary source
- October 29 FP8/NVFP4 release ↗
Hugging Face · Primary source