Le Chonk AI (Mistral Large 4)

Le Chonk is Mistral Large 4, a multimodal AI model for developers building coding agents, document analysis, and tool-based workflows.

Le Chonk and Mistral Large 4 are the same model

Mistral introduced the public preview on October 6, 2026. Le Chonk is its nickname; ML4 and Mistral Large 4 refer to the same release. The practical starting point is the hosted preview in Mistral Studio, where developers can evaluate it before committing to an integration. The model identifier in the official documentation is mistral-large-4.

This is a model to connect to an application or agent. A working agent still needs its tools, permissions, and execution environment. For example, a document-analysis application can supply a technical drawing and ask the model to locate a component, then let a tool zoom in for a closer inspection. A coding agent can supply repository context and expose a terminal for the model to propose and test changes. Those workflows explain the release's focus on visual understanding and multi-step tool use.

What developers can build with the preview

• Repository and terminal workflows: Mistral evaluates ML4 on software changes, repository questions, and tasks involving command-line tools. Teams can test it against their own code and existing agent harness.
• Image and document analysis: the announcement demonstrates locating objects in dense scenes, inspecting mechanical drawings, retrieving evidence from PDFs, and searching large geospatial images.
• Structured application output: the model card lists structured outputs and function calling, useful when an application needs fields or tool arguments rather than a free-form answer.
• Business workflows: the release discusses agents that gather information across applications and produce spreadsheets, slides, and documents. Access to those applications comes from the surrounding integration.
• Scientific programming: Mistral presents data analysis, modeling, and simulation as use cases, including a chemistry simulation example. Generated code and scientific conclusions still need task-specific verification.

The model card lists a one-million-token context window and support for chat completions, document questions, batching, Agents and Conversations, and built-in tools. Check the endpoint you intend to use: a listed model capability and the behavior of a particular application are different things.

API access now, downloadable weights later

As of October 7, the API is in public preview. Mistral's English release announcement says weights will arrive by the end of October, while its changelog says open weights are coming soon. Downloadable weights and local deployment are therefore future release milestones, even though the product is described as an open-weight model.

For teams considering private hosting, wait for the actual checkpoints, license, deployment instructions, and hardware requirements before budgeting an installation. The announcement describes a future private-cloud or on-premise route for organizations that need control over their security workflows; it does not establish that the preview is a lightweight model you can run on a laptop.

The official sources also differ on the architecture numbers. The announcement describes approximately one trillion total parameters and 49 billion active parameters; the model card lists 1.05 trillion total and 52 billion active, plus a vision encoder. Both describe a very large multimodal model. Use the eventual released checkpoint documentation for deployment specifications rather than treating either preview figure as a final requirement.

What the European infrastructure claim means

Mistral says ML4 was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters and that the public preview runs on that infrastructure. Its planned regional offerings include a European deployment operated end-to-end by Mistral under European law. For an enterprise choosing a provider, this is a concrete infrastructure and control proposition, separate from whether a model wins a benchmark.

The company is using the preview period for real-world evaluations with cybersecurity partners and public authorities. It also reports multilingual training data spanning more than 160 languages. These are Mistral's descriptions of its development and evaluation process; they do not promise equal accuracy in every language or unrestricted access to every partner configuration.

Evaluate the task, then estimate the cost

The release highlights coding, cybersecurity, finance, legal work, and visual grounding. Its benchmark results are reported by Mistral, including evaluations conducted with outside organizations. They are useful starting points for choosing tests, rather than a guarantee that ML4 will outperform another model on your particular codebase or documents. We have not run an independent generation or agent evaluation for this listing.

The published standard API rates are $1.36 per million input tokens, $0.14 for cached input, and $4.18 for output. The model card displays launch rates of $0.68, $0.07, and $2.09 respectively; the changelog describes a 50% launch discount for two weeks. Treat those lower prices as temporary and check the current provider quote before estimating ongoing usage.

Start with a small set of representative tasks in Studio, inspect the answers and tool calls, and compare total token use as well as output quality. Long agent runs may require repeated calls and external tools. The official image used on this page is a release graphic, not a screenshot of an independently tested application.