Mistral Large 4 is an open-weight, multimodal Mixture-of-Experts large language model from Mistral AI, released as a public preview on 6 October 2026 and listed as generally available on its model card from 5 November 2026. This profile collects the published facts about it in one place: API identifiers, context window, parameter counts, prices, reasoning modes, licence terms, where it can be run, provider-reported benchmarks, data handling and deprecation status. It is written for developers, architects, buyers and governance teams who want checked numbers rather than launch coverage. Every fact comes from Mistral AI's own documentation, model card, launch post, Hugging Face model card, pricing page or help centre, listed under Standards and references. Where Mistral AI publishes nothing on a point, the page says so.
Last checked: 2026-11-15.
This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider. It uses Mistral AI's public documentation as of November 2026. Model details change, so confirm against Mistral AI's current documentation.
Key takeaways
- Preview 6 October 2026, GA 5 November 2026. The Mistral changelog records the public preview of mistral-large-4 on 6 October 2026; the model card gives version 26.10, a release date of 5 November 2026 and status GA.
- 1M token context, 1.05T total and 52B active parameters. The model card describes a granular Mixture-of-Experts design with a 1.6B-parameter vision encoder; the Hugging Face card gives 128 routed experts with 4 active per token plus 1 shared expert.
- $0.70 input, $0.07 cached input and $2.10 output per million tokens. These are the model card and launch post figures. Mistral's general pricing page FAQ quotes "Mistral Large" at $0.5 in and $1.5 out, so confirm the billed rate in the console.
- Open weights under the Mistral Ecosystem Open-Weight License. Weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B; the licence is described as a modified MIT licence, with a separate commercial licence required to run a Model-as-a-Service business.
- No deprecation date. Mistral's models overview lists Large 4 as current, and it is the named replacement for Mistral Large 3, which retires on 30 April 2027.
What is Mistral Large 4?
Mistral Large 4 is the current flagship of Mistral AI's Large line. The Hugging Face model card calls it "our largest and most capable model to date", and the launch post describes a 1-trillion-parameter, natively multimodal model trained from scratch on European infrastructure, nicknamed "Le Chonk". The Hugging Face card states that it was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacentres.
It accepts text and images and returns text, with an optional reasoning trace. Mistral's model card lists the output modalities as "reasoning, text". The launch post says the weights followed the public preview after about three weeks of additional reinforcement learning, and that Large 4 will serve as the foundation for future specialised Mistral models. Because only part of the network is active per token, Mistral gives both a total and an active parameter count; see Mixture of Experts Explained. For the full Mistral line-up, see Mistral AI Models.
Key facts at a glance
As of 2026-11-15. Sources: Mistral Large 4 model card (docs.mistral.ai), Hugging Face model card, Mistral changelog and launch post.
| Field | Value | Source |
|---|---|---|
| Provider | Mistral AI | Model card |
| API ID | mistral-large-4, mistral-large-4-0, mistral-large-latest | Model card |
| Version | 26.10 | Model card; models overview |
| Release date | Public preview 2026-10-06; model card release date 2026-11-05 | Changelog; model card |
| Status | GA, current | Model card; models overview |
| Context window | 1M tokens | Model card; Hugging Face card |
| Max output | Not stated by Mistral AI on the pages checked (2026-11-15) | Model card |
| Knowledge cutoff | Not stated by Mistral AI on the pages checked (2026-11-15) | Model card; Hugging Face card |
| Input modalities | Text, image | Model card |
| Output modalities | Reasoning, text | Model card |
| Reasoning effort | low, medium, high (default) | Hugging Face card; launch post |
| Open weights | Yes, mistralai/Mistral-Large-4-1T-A52B | Model card; Hugging Face card |
| Licence | Mistral Ecosystem Open-Weight License (modified MIT) | Model card; Hugging Face card |
| Parameters | 1.05T total, 52B active (49B excluding embeddings and output layers), plus 1.6B vision encoder | Model card; Hugging Face card |
| Languages | 160+ languages, including all official EU languages | Hugging Face card |
Release and version history
As of 2026-11-15. Dates from the Mistral changelog, models overview and model cards.
| Date | Event | Source |
|---|---|---|
| 2024-02-26 | mistral-large-latest first introduced, then pointing to mistral-large-2402 (Large 1.0) | Changelog |
| 2025-12-02 | Mistral Large 3 (mistral-large-2512) released | Changelog; models overview |
| 2026-10-06 | Mistral Large 4 (mistral-large-4) in public preview; open weights "coming soon"; launch pricing 50% off for two weeks | Changelog |
| 2026-10-31 | Mistral Large 3 deprecated, with Large 4 named as replacement | Models overview |
| 2026-11-05 | Model card release date, status GA; weights published on Hugging Face | Model card; launch post; Hugging Face card |
| 2027-04-30 | Scheduled retirement of Mistral Large 3 | Models overview |
The launch post reports how the release changed from the preview on several evaluations (see Benchmarks). Mistral publishes one version, 26.10, with three API names. The floating alias mistral-large-latest has pointed to different Large releases over time, so code that needs stable behaviour should call mistral-large-4-0 instead. See LLM Naming and Versioning Explained.
How much does Mistral Large 4 cost?
As of 2026-11-15. USD per million tokens. Sources: Mistral Large 4 model card, launch post and mistral.ai/pricing.
| Item | Price | Notes | Source |
|---|---|---|---|
| Input | $0.70 | Model card also offers a EUR view | Model card; launch post |
| Cached input | $0.07 | 10% of input rate | Model card; launch post |
| Output | $2.10 | Reasoning tokens are generated as output | Model card; launch post |
| Batch | 50% reduction | General statement on the pricing page, not model specific | mistral.ai/pricing |
| "Mistral Large" (unversioned) | $0.5 input, $1.5 output | Pricing page FAQ; differs from the model card | mistral.ai/pricing |
| Self-hosted weights | No per-token fee from Mistral; own compute cost | Model-as-a-Service use needs a separate commercial licence | Hugging Face card; launch post |
Price conflict: two official Mistral pages give different figures for "Large". The model card is the page specific to mistral-large-4, and the launch post repeats its figures, so this profile uses $0.70, $0.07 and $2.10. Check the rate shown in Mistral Studio before committing a budget.
Worked example at model card rates: a request with 100,000 input tokens and 10,000 output tokens costs $0.07 for input plus $0.021 for output, about $0.091 in total. At high reasoning effort the output count includes reasoning tokens, so measure real token use per task. See Model Pricing Explained and How to set spend limits on LLM API keys.
Context window and output limits
The context window is 1M tokens on both the Mistral model card and the Hugging Face card. Mistral does not publish a separate maximum output figure on the pages checked. The Hugging Face card notes that when serving with vLLM, the max_completion_tokens setting counts reasoning tokens as well as visible output, so a low cap can cut off a high-effort answer.
| Limit | Value | Notes | Source |
|---|---|---|---|
| Context window | 1M tokens | Same figure on API and open weights | Model card; Hugging Face card |
| Max output | Not stated by Mistral AI on the pages checked (2026-11-15) | None | Model card |
| Reasoning tokens | Counted within completion tokens | vLLM max_completion_tokens | Hugging Face card |
| API rate limits | Not stated by Mistral AI on the pages checked (2026-11-15) | Set per workspace in Mistral Studio | Model card |
| Self-hosting example | Tensor parallel size 8 in the published vLLM example | NVFP4 checkpoint targets Blackwell GPUs | Hugging Face card |
Modalities and tools
Large 4 takes text and images and returns text. The Hugging Face card says its Pixtral-style vision encoder handles variable image sizes and aspect ratios, and lists images, charts, technical drawings and documents as vision use cases.
- API features listed on the model card: chat completions, function calling, structured outputs, document Q&A, prefix, batching, agents and conversations, and connectors.
- Reasoning effort: low, medium and high (default), set per request. The launch post says low replaces the former "None" mode; Mistral suggests high for complex or agentic work, medium for everyday use and low for quick chat.
- Recommended sampling: temperature 1.0 and top_p 0.95 for all reasoning efforts (Hugging Face card).
- Tool use: native function calling, JSON output and strict JSON schema enforcement. Mistral advises keeping the tool set well defined and as small as the use case allows, and stating the environment and tool guidance in the system prompt.
- Self-hosting stacks named by Mistral: vLLM (recommended), SGLang, Transformers, llama.cpp (GGUF) and Docker Model Runner, with fine-tuning through Axolotl, Unsloth and TRL. An embedded draft head supports speculative decoding, which the card reports gives up to 1.9x higher throughput depending on hardware and workload.
Connectors and agents give the model reach into other systems. Treat each connected tool as a credentialed integration with its own least-privilege scope. See the Open-Weight Model Security Guide for self-hosted deployments.
Where is Mistral Large 4 available?
As of 2026-11-15. Channels as named in the launch post, the model card and the Hugging Face card.
| Channel | Details | Source |
|---|---|---|
| Mistral API and Mistral Studio | mistral-large-4, mistral-large-4-0, mistral-large-latest | Model card; launch post |
| Hugging Face weights | mistralai/Mistral-Large-4-1T-A52B, plus an NVFP4 quantised checkpoint | Hugging Face card |
| Microsoft Foundry | Listed; the launch post says Microsoft describes itself as the first cloud provider to offer it, with private deployment options for Azure | Launch post |
| NVIDIA | NVIDIA-hosted endpoints, an NVIDIA NIM container, NeMo AutoModel and TensorRT-LLM support | Launch post |
| OpenRouter | Listed | Launch post |
| Vibe Code and Mistral Vibe CLI | Listed as able to use the model | Launch post; Hugging Face card |
| Mistral Forge | Listed for customisation | Launch post |
| Amazon Bedrock, Google Vertex AI | Not stated by Mistral AI on the pages checked (2026-11-15) | Launch post |
Access restrictions
The general Large 4 model is not restricted: it is GA on the Mistral API and its weights are public on Hugging Face. The licence places one commercial condition on self-hosting: a business that offers the model as a Model-as-a-Service needs a separate commercial licence from Mistral. The launch post also describes a specialised cybersecurity variant that is in preview only for vetted enterprise design partners, such as Orange Cyberdefense, and restricted to owned or authorised systems. That variant is separate from the public model.
Benchmarks (provider-reported, dated)
As of 2026-11-15. All figures are as reported by Mistral AI in the Large 4 launch post or Hugging Face model card (November 2026). Figures attributed to Artificial Analysis are as quoted by Mistral; they were not checked against Artificial Analysis's own pages for this profile.
| Evaluation | Mistral Large 4 | Comparator in source | Source and date |
|---|---|---|---|
| Artificial Analysis Coding Agent Index | 48% | "Effectively tied" with GLM-5.3, with 47% fewer tokens per task on average | Launch post; Hugging Face card |
| Artificial Analysis Intelligence Index v5 | 26 | Mistral states leading Chinese open-weight models score higher | Hugging Face card |
| Artificial Analysis Cyber Index | 51% | None given | Launch post |
| Terminal-Bench | 41% | Preview: 28%, with 30% fewer output tokens in the release | Launch post |
| Cognition FrontierCode 2.0 (private) | 19.7% | Top score shown in the post | Launch post |
| FinWorkBench (Finch) | 73.3% | Preview: 67.4% | Launch post |
| BrowseComp | 84.5% | Earlier SFT checkpoint: 75% | Launch post |
| B3 Agent Security Benchmark (attack resistance) | 97.1% | Preview: 93.3% | Launch post |
| CWE-bench resolve rate | 46.3 | None given | Hugging Face card |
| Malicious cyber request refusal rate | 96.6% average | Mistral says highest it measured among open models | Launch post |
The launch post also reports a Surge Code blind human preference result in which Large 4 was preferred over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash and placed behind Claude Opus 5.5. Benchmark versions, harnesses and settings differ between labs, so treat cross-lab comparisons with care.
Model card and system card
Mistral publishes two model cards for Large 4: the documentation card at docs.mistral.ai, with API names, pricing, features and licence, and the Hugging Face card, with architecture, recommended settings and deployment guidance. The Hugging Face repository also holds a technical report (assets/technical_report.pdf). Mistral does not publish a separate system card for Large 4 on the pages checked. Safety statements come from the launch post:
- Cyber guardrails: built-in guardrails are intended to distinguish defensive from malicious cybersecurity tasks, with a 96.6% average refusal rate on malicious cyber requests.
- Residual risk: Mistral says it observed refusals to black-box attacks on live targets but sees potential for uplift to real attacks, and that it is modifying its guardrails.
- Restricted variant: the specialised cybersecurity variant is limited to vetted design partners.
- Licence wording: the launch post's model box shows "Apache 2.0", while its text, the documentation card and the Hugging Face card name the Mistral Ecosystem Open-Weight License. The documentation card is used here.
For what model and system cards usually contain, see Model Cards and System Cards Explained.
Data handling
Mistral AI's help centre applies the same rules to Large 4 as to other API models; there is no Large 4 specific data policy on the pages checked.
- Free mode in Mistral Studio: Mistral states it may use inputs and outputs to train its models, with an opt-out available.
- Pay-as-you-go: Mistral states customers keep control over this processing and can opt out at any time.
- Zero Data Retention: available only on pay-as-you-go and only for stateless API calls (chat completions, FIM, embeddings, moderation and classification, OCR, audio). It is granted on request, at Mistral's discretion.
- ZDR exclusions: agents, batch files, conversations, libraries, the Files API, Vibe Work and Chat. Large 4's agents, conversations, connectors and batching features therefore sit outside ZDR.
- Self-hosted weights: data stays wherever you run the model; retention is your own responsibility.
Mistral API keys are non-human identities with direct spend and data exposure. See Zero Data Retention for AI and the LLMjacking Guide.
Deprecation status
As of 15 November 2026, Mistral's models overview lists Mistral Large 4 as a current model with no deprecation or retirement date. It is the named replacement for Mistral Large 3, and the models overview also names it as the alternative for Mistral Medium 3.5 (deprecated 2026-10-31, retiring 2027-04-30). Watch the models overview and changelog for changes.
| Model or item | Details | Source |
|---|---|---|
| Mistral Large 4 (26.10) | Current; no deprecation or retirement date listed (2026-11-15) | Models overview |
| Mistral Large 3 (mistral-large-2512) | Deprecated 2026-10-31, retires 2027-04-30; replacement Mistral Large 4 | Models overview |
| mistral-large-latest alias | Points to Large 4 per the model card; has pointed to earlier Large releases before | Model card; changelog |
How Mistral Large 4 differs from other models in its family
As of 2026-11-15. Sources: Mistral model cards, models overview and changelog. Models listed by release date.
| Model | Released | Context | Parameters | Licence | Status |
|---|---|---|---|---|---|
| Mistral Large 3 | 2025-12-02 | Not stated on the pages checked | Not stated on the pages checked | Not stated on the pages checked | Deprecated, retires 2027-04-30 |
| Mistral Small 4 | 2026-03-16 | 256k | 119B total, 6.5B active | Apache 2.0 | GA |
| Mistral Medium 3.5 | Version 26.04 | Not stated on the pages checked | Not stated by Mistral AI | Not stated on the pages checked | Deprecated, retires 2027-04-30 |
| Mistral Large 4 | Preview 2026-10-06; GA 2026-11-05 | 1M | 1.05T total, 52B active | Mistral Ecosystem Open-Weight License | GA |
The documented differences from Small 4 are context (1M versus 256k), scale, reasoning modes (Large 4 offers low, medium and high; Small 4 offers none and high), licence terms, and the connectors feature listed on the Large 4 card. See Mistral Small 4, Reasoning Models Explained and Open-Weight Frontier Models Compared.
Practitioner checklist
- Pin mistral-large-4-0 in production code rather than mistral-large-latest, and record the date you adopted it.
- Confirm the billed price in Mistral Studio, given the two official figures, and model cost with reasoning tokens counted as output.
- Pick a reasoning effort per workload (low, medium or high) and measure tokens per task; high is the default.
- Read the Mistral Ecosystem Open-Weight License before self-hosting, and get a commercial licence if you will offer the model as a service.
- Check whether your use relies on agents, conversations, batch or connectors, which sit outside Zero Data Retention.
- Scope connectors and function-calling tools to least privilege, with approval for destructive actions.
- Store Mistral API keys in a secrets manager, separate workspaces by environment and set spend limits.
- If you are migrating from Mistral Large 3, plan the move before its 30 April 2027 retirement and retest prompts.
Standards and references
- Mistral AI: Mistral Large 4 model card (accessed 2026-11-15)
- Mistral AI on Hugging Face: Mistral-Large-4-1T-A52B model card (accessed 2026-11-15)
- Mistral AI: Mistral Large 4 launch post (accessed 2026-11-15)
- Mistral AI: Changelog (accessed 2026-11-15)
- Mistral AI: Models overview (accessed 2026-11-15)
- Mistral AI: Pricing (accessed 2026-11-15)
- Mistral AI Help Center: Can I activate Zero Data Retention (ZDR)? (accessed 2026-11-15)
- Mistral AI Help Center: Do you use my user data to train your models? (accessed 2026-11-15)
Related NHI Mgmt Group resources: Mistral AI Models, Mistral Small 4, Open-Weight Frontier Models Compared, Open-Weight Model Licence Tracker, LLM Pricing Tracker.