Mistral Large 4: API Pricing, Specs, Context and Open Weights
Mistral Large 4 is Mistralโs new flagship model for multimodal reasoning, coding and tool-driven workflows. The practical distinction is availability: the API is GA and official downloadable weights are available. This guide separates what a developer can buy today from the separately licensed self-hosted route.
Availability and model identity
Mistral announced Large 4 on October 6, 2026 and released improved weights on October 29. The current model documentation labels the API GA, using mistral-large-4. Re-evaluate the released checkpoint separately from the earlier preview.
| Item | Verified current state on November 7 |
|---|---|
| API | GA, usable now |
| Model ID | mistral-large-4 |
| Context | 1 million tokens in the model documentation |
| Inputs | Text and images; native multimodal architecture |
| Outputs | Text |
| Architecture | Mixture of experts; approximately 1 trillion total parameters |
| Active parameters | 52B including embeddings and output layers; 49B excluding them, per the released model card |
| Weights | Available from Mistral on Hugging Face |
Active parameters: 52B, and why you may still see 49B
The released model card now explains the counting difference: 49B active per token, or 52B when embeddings and output layers are included. The API documentation lists 1.05T total parameters and a 1.6B vision encoder. Preserve the counting convention when comparing models. Active parameters do not represent the resident weight memory.
API pricing: current rate, not the expired preview offer
The current official model card, checked November 7, displays $0.70 input, $0.07 cached input and $2.10 output per million tokens. These are the displayed direct USD rates. The original two-week, 50% preview offer is historical; do not continue quoting its $0.68/$2.09 figures or treat the old displayed list rate as todayโs rate.
One million uncached input tokens plus 100,000 output tokens costs $0.91 at the current displayed rates, before other applicable charges. This is arithmetic, not an application benchmark. Cached-input billing requires a qualifying cache hit. Recheck the selected route when budgeting; third-party hosting and subscriptions can differ.
Capabilities and integration choices
The official card lists structured outputs, function calling and document question answering, including chat-completions and conversations interfaces. That makes the model relevant to coding assistants, document analysis and agents that must return constrained data or invoke tools. A million-token context is a capacity specification, not a guarantee of reliable retrieval across every token.
Start with an isolated evaluation against your current model: the same tasks, permissions, tool descriptions, error handling and cost accounting. Verify your SDK or gateway passes the documented model ID, inspect usage fields, and test tool-call failures as well as successful responses. Do not silently swap an existing production alias for a changing preview.
Benchmarks: promising, but not our test results
Mistralโs launch material emphasizes agentic coding, business workflows and visual understanding. These are vendor-reported evaluations, not AI Made Tools hands-on benchmarks. Their task sets, scaffolds, refusal policies and scoring differ, so they cannot establish a universal winner or be mixed into an older SWE-bench or HumanEval ranking. We have not independently measured Large 4โs coding reliability, latency or prompt-injection resistance.
For a useful buying decision, evaluate repository tasks, long documents or tool workflows representative of your own application. Compare completed-task cost, not token price alone. Our agent model guide explains the operational dimensions; the Mistral model directory helps distinguish the flagship model from smaller and older models.
Open weights and local deployment
The official checkpoint and an NVFP4 variant are downloadable. The model card documents vLLM, SGLang and llama.cpp routes. Use the checkpointโs current recipe and pin revisions rather than assuming an arbitrary installed runtime supports this architecture. The vLLM example uses eight-way tensor parallelism; that example is not a verified minimum GPU budget.
The Mistral Ecosystem Open-Weight License is a modified license, not an unrestricted MIT release. The card permits commercial and non-commercial use but requires a separate commercial license for operating a Model-as-a-Service business. Check the actual license for your distribution and serving arrangement.
The released model supports low, medium and default high reasoning effort. Reasoning consumes part of the completion budget. Test reasoning and tool parsers with the selected runtime. We have not measured self-hosted throughput, maximum usable context or GPU memory consumption.
Should you use it now?
- Evaluate the released model if native vision, document reasoning or multi-step tool use matters.
- Keep a validated production route until the replacement passes your task and permission checks.
- Evaluate self-hosting only after reviewing the license, hardware footprint and supported runtime recipe. Downloadability does not imply workstation practicality.
This is a documentation-based developer reference. We have not independently benchmarked the released checkpoint.