Quick Answer
Mistral Large 4 open weights are now downloadable under a license that permits many commercial uses, modification, and fine-tuning. Mistral released the improved model on October 29, 2026, with a 1 million-token context window and new coding and cybersecurity claims. Developers should review the license before hosting the model as a commercial service and validate vendor-reported benchmarks in their own workloads.
Key Takeaways
- Mistral released the improved Large 4 weights on October 29, 2026, after an earlier public preview.
- Mistral Large 4 is a 1-trillion-parameter mixture-of-experts model with 52 billion active parameters.
- The model accepts text and images, produces text, and supports a 1 million-token context window.
- Mistral reports new coding and cybersecurity benchmark results, but those claims remain vendor-reported.
- The Ecosystem Open-Weight License permits many commercial uses but restricts commercial model-as-a-service operations.
What did Mistral release for Mistral Large 4?
Mistral released downloadable weights for the improved Mistral Large 4 model on October 29, 2026. The release follows the model’s public preview earlier in October and gives developers a way to run, modify, and fine-tune the model outside a purely hosted API workflow. Mistral’s open-weights announcement describes the release as an updated version of Large 4, also called “Le Chonk.”
Mistral Large 4 is a natively multimodal mixture-of-experts model. A mixture-of-experts design activates part of a larger model for a given task rather than using every parameter in the same way for each request. The practical benefit is that developers can evaluate a large model architecture while working with a stated active-parameter count that is lower than its total parameter count.
The release matters most for organizations that need more control over model deployment, evaluation, and fine-tuning. Open weights do not remove all licensing obligations, security responsibilities, or infrastructure costs. Developers comparing locally controlled models with hosted products should also consider how free AI tools handle data, access controls, and feature limits before moving a production workflow to any new model.
What are Mistral Large 4’s core specifications?
Mistral Large 4 has 1 trillion total parameters and 52 billion active parameters, including embeddings and output layers. The official model card identifies the model as natively multimodal and states that it accepts text and image input while producing text output. Mistral’s Hugging Face model card provides the published model details and downloadable weights.
Mistral Large 4 also supports a 1 million-token context window, function calling, strict JSON schema output, and low, medium, and high reasoning settings. A context window is the amount of input material a model can consider in one request, while strict JSON schema output is intended to make generated responses fit a required structured format. Those features can help with large document collections, tool-connected workflows, and applications that require predictable machine-readable output.
The specification list does not establish how well the model will perform on a particular company’s documents, codebase, or image inputs. Prompt design, retrieval setup, tool permissions, and evaluation methods still affect results. The sensible approach is to test a small set of representative tasks before treating the published context limit or feature list as proof of production readiness.
| Published capability | What Mistral lists | Why it matters |
|---|---|---|
| Model architecture | 1 trillion parameters, 52 billion active parameters | The model uses a mixture-of-experts approach rather than activating its full parameter total for every task. |
| Inputs and output | Text and image input, text output | Applications can send written material and images into one model workflow. |
| Context window | 1 million tokens | Developers can assess large source collections or long task histories in a single request. |
| Structured workflows | Function calling and strict JSON schema output | Software can request tool actions and more predictable structured responses. |
| Reasoning controls | Low, medium, and high settings | Developers can choose among the available reasoning modes for different tasks. |
What coding results does Mistral claim for Large 4?
Mistral reports that Large 4 scored 19.7% on Cognition’s private FrontierCode 2.0 benchmark and 48% on Artificial Analysis’ Coding Agent Index. Mistral also says Large 4 used 47% fewer tokens per task on average than GLM-5.3 while tying that model on the Artificial Analysis Coding Agent Index. These are company-reported benchmark claims, not independently verified performance conclusions.
The coding claims matter because token consumption affects the cost and speed of an agentic coding workflow. A model that completes comparable tasks with fewer tokens could reduce API usage, but benchmark outcomes do not guarantee the same result on a private repository or a specific engineering process. Tool access, task design, programming language, test coverage, and human review can materially change an AI coding system’s usefulness.
Mistral Large 4’s function calling and JSON schema support could be relevant for coding agents that need to call development tools and return structured results. At the same time, developers should not allow a model to make unrestricted changes to production systems based on benchmark scores alone. Teams using AI for code analysis should keep review gates, testing requirements, and limited credentials in place, particularly where the model can invoke external tools.
What are Mistral Large 4’s cybersecurity claims?
Mistral reports that Large 4 scored 51% on the Artificial Analysis Cyber Index and achieved an average 96.6% refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm prompts. Mistral presents those results as safety and cybersecurity evaluations, but they remain vendor-reported results that require independent validation before organizations rely on them for risk decisions.
Cybersecurity benchmark results measure specific test conditions rather than every real-world misuse path. A high refusal rate can indicate that a model declines many harmful requests, but a refusal metric does not replace access controls, human review, logging, or incident response. The most important point is that a model’s safety behavior depends on the surrounding application as well as the base model.
Organizations that connect Mistral Large 4 to security tools should apply least-privilege permissions and isolate test environments before granting access to sensitive systems. The same principle applies to AI systems that identify possible exploit chains, because a model output can be incomplete or wrong even when it sounds confident. Teams reviewing AI-assisted exploit analysis should treat model results as input for qualified security review, not as automated authorization to act.
What license applies to Mistral Large 4 open weights?
Mistral Large 4 uses Mistral’s Ecosystem Open-Weight License, a modified MIT-style license that permits commercial use, modification, and fine-tuning. The license requires a separate agreement for commercial model-as-a-service operators. That distinction is important because downloading open weights does not automatically grant every type of commercial hosting right.
The license structure gives many developers room to adapt the model for internal tools, products, and research. A company that plans to expose the model as a paid or commercial service must determine whether its planned use falls under the separate licensing requirement. Legal terms can affect product architecture, vendor selection, and launch timing, so the license review belongs before deployment rather than after a service reaches users.
Mistral Large 4’s release also shows why open-weight availability and unrestricted use are not the same thing. Organizations should read the current license text, document how the model will be distributed or hosted, and involve qualified legal counsel when the business model depends on interpretation of a license condition. The practical response is to treat licensing as a deployment requirement alongside security testing and data governance.
How much does Mistral Large 4 cost through the API?
Mistral lists API pricing for Large 4 at $0.70 per million input tokens, $0.07 per million cached-input tokens, and $2.10 per million output tokens. Mistral’s Large 4 documentation lists those rates and the model’s published capabilities. API pricing can change, so developers should confirm the current rate before estimating a production budget.
Input tokens are the text or other tokenized material sent to the model, while output tokens are the model’s generated response. Cached-input pricing can matter for workflows that repeatedly reuse the same material, such as a stable system instruction or a large recurring reference document. The lower cached-input rate does not mean every request qualifies for that rate, so usage design still determines the final bill.
The pricing structure does not include the cost of application hosting, storage, monitoring, retrieval systems, human review, or security controls. Organizations deciding between self-hosting the weights and using the API should calculate the full operating cost for their own request volumes and governance needs. Users comparing connected AI services should also understand that AI access to more apps can introduce additional data handling considerations beyond token pricing.
Who should consider Mistral Large 4 open weights?
Mistral Large 4 open weights are most relevant to developers and organizations that need to evaluate, customize, fine-tune, or control the deployment of a multimodal model. The 1 million-token context window, image input, function calling, and structured-output features give technical teams several areas to test. The model is less suitable for users who only need a simple consumer chatbot and do not have a reason to manage model infrastructure or licensing terms.
Developers should begin with a limited evaluation that measures the model against tasks that reflect their actual requirements. Useful tests can include structured extraction, document analysis, coding tasks, image-based prompts, and tool-call reliability. A benchmark score is not a substitute for a controlled internal test, particularly when an application handles confidential material or can take actions through connected systems.
- Review the Mistral Large 4 license and determine whether the planned deployment involves a commercial model-as-a-service offering.
- Test the model on representative prompts, documents, images, and coding tasks before integrating it into a broader workflow.
- Restrict tool permissions and keep sensitive data out of early experiments unless appropriate controls are in place.
- Compare token use, output quality, and failure cases across the available reasoning settings.
- Stop and consult legal, security, or vendor support teams before launching a commercial hosted service or granting the model access to sensitive production systems.
Mistral Large 4’s downloadable weights provide flexibility, but flexibility also shifts more responsibility to the deployer. Organizations should establish review processes for model updates, user data, tool permissions, and safety testing. Readers considering how model training and deployment affect data rights may also want to understand the emerging shared AI training opt-out approach, especially where proprietary content is involved.
FAQ
Are Mistral Large 4 open weights free to use commercially?
Mistral Large 4 open weights permit commercial use, modification, and fine-tuning under Mistral’s Ecosystem Open-Weight License. Commercial model-as-a-service operators require a separate license, so hosted-service plans need a license review.
How large is Mistral Large 4?
Mistral Large 4 has 1 trillion total parameters and 52 billion active parameters, including embeddings and output layers. The model uses a mixture-of-experts architecture, which activates a subset of the overall model for a task.
Does Mistral Large 4 support image input?
Mistral Large 4 supports text and image input and produces text output. The model also supports function calling, strict JSON schema output, and a 1 million-token context window.
Are Mistral’s Large 4 cybersecurity scores independently verified?
No, Mistral’s Large 4 cybersecurity scores are company-reported claims. Organizations should validate safety behavior in their own environment and retain access controls, monitoring, and human review.
What does Mistral Large 4 cost through the API?
Mistral lists Large 4 API pricing at $0.70 per million input tokens, $0.07 per million cached-input tokens, and $2.10 per million output tokens. Developers should confirm current pricing before deployment because API rates can change.