Quick Answer
Mistral Large 4 is a public-preview multimodal AI model with 1.05 trillion total parameters and a 1 million-token context window. Mistral offers the model through its moderated API, while planned open weights remain under safety testing. Developers can evaluate its document, agent, and tool-use features now, but should treat performance and availability claims as preview-stage information.
Key Takeaways
- Mistral Large 4 has 1.05 trillion total parameters and 49 billion active parameters.
- The model accepts multimodal inputs and includes a 1.6 billion-parameter vision encoder.
- Mistral Large 4 supports a 1 million-token context window, function calling, and structured outputs.
- Public-preview API pricing starts at $0.68 per million input tokens in the US-facing documentation.
- Mistral plans to release the model weights after further red-team testing in October.
What is Mistral Large 4?
Mistral Large 4 is Mistral’s new public-preview large language model, released on October 6, 2026, with native support for text and image inputs. Mistral calls the model “Le Chonk” and has made API access available through Mistral Studio. Mistral’s announcement describes Large 4 as its newest frontier model for production AI work.
Mistral Large 4 uses a mixture-of-experts design, which means the full model contains a large pool of parameters while each request activates only part of that pool. Mistral lists 1.05 trillion total parameters and 49 billion active parameters. The distinction matters because total parameter counts alone do not show the computing cost or speed of a model during normal use.
Mistral Large 4 is not yet a fully open-weight release. The current public-preview phase gives developers API access while Mistral continues safety work, which limits what can be concluded about long-term availability, final safeguards, and independent performance. Developers evaluating the model should begin with bounded tests that use representative documents, tools, and workflow prompts rather than moving sensitive production tasks immediately.
What technical specifications does Mistral Large 4 have?
Mistral Large 4 has a 1 million-token context window, native multimodal capabilities, and a 1.6 billion-parameter vision encoder. Mistral’s model documentation also lists support for structured outputs, function calling, document question answering, chat completions, batching, agents, and conversations.
A 1 million-token context window allows a developer to supply unusually large amounts of text in a single request, such as long technical documentation, a substantial codebase, or a large collection of business records. The practical benefit is that an application may need fewer separate retrieval and summarization steps. The limitation is that a large context window does not establish that every detail receives equal attention or that every answer is correct.
| Specification | Mistral Large 4 detail | Why it matters |
|---|---|---|
| Model architecture | Native multimodal mixture of experts | The model can process text and image inputs while activating 49 billion parameters per request. |
| Total parameters | 1.05 trillion | The total figure describes the model’s full parameter pool, not necessarily the cost of each inference. |
| Vision encoder | 1.6 billion parameters | The vision component supports image-related inputs within the wider model system. |
| Context window | 1 million tokens | Applications can submit very large source sets in a single context. |
| Tool features | Structured outputs and function calling | Developers can connect model responses to software workflows in more controlled formats. |
Mistral Large 4’s specification list is useful for planning integrations, but it is not an independent quality ranking. Developers should test the model against their own error tolerance, document types, image quality, tool permissions, and response-format requirements. Readers comparing major model releases can also review the different product approach in Claude Sonnet 5.5’s launch.
What can developers do with Mistral Large 4?
Mistral Large 4 can support document question answering, chat applications, agent workflows, batching, structured outputs, and function calling through its API. Structured outputs are responses constrained to a requested format, while function calling allows an application to pass a model-selected action to external software. These capabilities can make model output easier for a program to validate and use.
Document question answering is one of the clearest uses for the model’s large context capacity. An organization could provide a large set of internal materials and ask for an answer grounded in those materials. The safety and accuracy limit remains important: a model can still misunderstand the source text, omit qualifying details, or produce an incorrect response that appears confident.
Agent features require more caution because a model connected to tools can affect systems beyond a chat window. Developers should limit tool permissions, require confirmation for consequential actions, log requests and outputs, and test failure paths before allowing automated tasks. The broader policy debate around AI systems and consumer risks has also reached government bodies, including the FTC’s AI agent investigation.
How much does Mistral Large 4 cost through the API?
Mistral Large 4 public-preview pricing is $0.68 per million input tokens, $0.07 per million cached-input tokens, and $2.09 per million output tokens. Mistral lists those rates in its model documentation as of October 6, 2026. Input tokens are the text or data sent to the model, while output tokens are the response generated by the model.
Cached-input pricing is lower because repeated context can be reused rather than processed from the beginning for every request. That difference can matter for applications that repeatedly reference the same long document collection or system instructions. Actual spending still depends on prompt length, response length, request volume, caching behavior, and any future changes Mistral makes during or after the preview.
Mistral Large 4 pricing should be evaluated against an application’s full operating cost, not only the published token rate. A workflow that sends large documents and requests extensive answers can generate substantial token use even when the input rate appears low. Developers should set budget limits, record token use during testing, and verify current pricing before committing to a production deployment.
How is Mistral handling safety testing for Large 4?
Mistral Large 4 is initially available through a moderated API while Mistral continues safety testing before the planned weight release. The company says it plans further red-teaming with cybersecurity leaders, vetted partners, and state authorities. That approach matters because broader access to powerful model weights can change how a model is tested, modified, and deployed outside the company’s own service.
Reuters reported that select cybersecurity experts and government authorities are receiving a version with fewer safety restrictions for testing. Reuters’ report also said Mistral CEO Arthur Mensch described Large 4 as stronger than Chinese models in some areas, including cybersecurity, without identifying the models or benchmarks.
The cybersecurity performance statement is therefore an attributed company claim, not a verified comparative ranking. The practical response for security teams is to assess the model through controlled, authorized testing and to keep proprietary code, credentials, customer data, and security-sensitive records out of early experiments unless contractual and technical protections are confirmed. Organizations should stop and consult their security, privacy, and legal teams before connecting a preview model to production systems.
When will Mistral release Large 4’s weights?
Mistral plans to release Mistral Large 4’s weights by the end of October after completing additional red-team work. Reuters reported that Mistral plans to make the model publicly available on October 27, 2026. The distinction between an API preview and released weights matters because API customers use Mistral-hosted access, while weights can enable deployment in other environments.
Open weights can give organizations more control over where and how a model runs, including possible self-hosted deployments. At the same time, weight availability places more responsibility on the organization deploying the model, especially for access controls, monitoring, data governance, and misuse prevention. The expected October timing remains a plan rather than a completed release.
Mistral Large 4 users should plan around the API that is available today rather than assume future self-hosting terms, hardware requirements, or final release conditions. Teams that need stable procurement details should wait for Mistral’s weight-release documentation before finalizing deployment architecture or compliance commitments.
What does Mistral Large 4 mean for the AI model market?
Mistral Large 4 gives developers another high-capacity multimodal model option with a 1 million-token context window and a stated plan for open weights. Axios reported that Mistral trained the model on 4,000 Nvidia Grace Blackwell GPUs over two months in European data centers. Axios’ reporting provides additional context on the training effort and Mistral’s moderated API rollout.
The launch does not establish that Mistral Large 4 is the best model for every workload. Model choice depends on the quality of outputs for a specific task, price, latency, data handling terms, tool support, deployment options, and reliability under real operational conditions. Vendor claims about comparative capability require independent testing before buyers treat them as procurement evidence.
For most developers, the useful next step is a limited evaluation against the models already used in their stack. Test the same prompt set, source documents, image inputs, tool calls, and response schemas, then review errors with subject-matter experts. Users handling personal information should apply the same caution recommended for AI chat privacy settings, because a model’s capability does not remove the need for data minimization and clear retention terms.
Who should evaluate Mistral Large 4 now?
Mistral Large 4 is most relevant now for developers who need large-context document analysis, multimodal inputs, structured responses, or tool-connected workflows and can test a public-preview API responsibly. The model’s documented features make it a plausible candidate for controlled pilots, especially where a long context window could reduce application complexity.
Organizations with strict requirements for stable pricing, independently verified benchmarks, final safety policies, or self-hosting should wait for more information. Mistral has not yet completed the planned weight release, and the research supplied for this launch does not establish independent accuracy, latency, or security results. A cautious evaluation should measure the model against the organization’s own requirements instead of relying on parameter counts alone.
Mistral Large 4 should not receive unrestricted access to sensitive systems during an early test. Start with non-sensitive sample data, restrict any functions the model can invoke, and review outputs before an action reaches a customer, employee record, financial process, or security environment. Security teams should set the stop line before deployment, not after an automated workflow causes an avoidable error.
FAQ
Is Mistral Large 4 available now?
Mistral Large 4 is available in public preview through Mistral Studio’s API as of October 6, 2026. Mistral is initially providing moderated API access while it continues safety testing before the planned weight release.
How many parameters does Mistral Large 4 have?
Mistral Large 4 has 1.05 trillion total parameters and 49 billion active parameters, according to Mistral’s documentation. The model also includes a 1.6 billion-parameter vision encoder for multimodal processing.
What is Mistral Large 4’s context window?
Mistral Large 4 has a 1 million-token context window. The large context capacity can support extensive document and code inputs, but users should still validate answers because more context does not guarantee complete accuracy.
How much does Mistral Large 4 cost?
Mistral Large 4 costs $0.68 per million input tokens, $0.07 per million cached-input tokens, and $2.09 per million output tokens during public preview. Actual costs depend on request volume, input length, output length, and caching use.
Will Mistral Large 4 have open weights?
Mistral Large 4 is planned for a weight release by the end of October after additional red-team testing. Reuters reported that Mistral expects public availability on October 27, 2026, but the final release terms and timing can change.