Mistral AI has unveiled Mistral Large 4, a multimodal model with 1.05 trillion parameters, and plans to make its weights available later this month. Nicknamed Le Chonk, the model entered public preview through Mistral’s API following its October 6 announcement, with pricing of $1.36 per million input tokens and $4.18 per million output tokens.

The release combines a mixture-of-experts architecture with a one-million-token context window. Mistral says the model supports more than 160 languages and is aimed at programming, cybersecurity and business workflows. Its planned downloadable weights would offer an alternative deployment route for organizations that want to operate the model themselves rather than rely exclusively on a hosted API.

Although Large 4 contains 1.05 trillion parameters in total, it activates 49 billion for each token it processes. Mixture-of-experts systems route work through selected parts of a model instead of using every parameter for every calculation. That design can reduce computation relative to a similarly sized dense model, but the total parameter count still has substantial implications for storing and hosting the weights.

The model also includes a 1.6-billion-parameter vision encoder, enabling it to process visual inputs alongside text. Its advertised context window sets the maximum amount of material it can handle within a single interaction. A large context capacity can be useful for lengthy documents and codebases, although capacity alone does not establish how reliably a model retrieves details or reasons across all of that material.

Public-preview access currently exposes two reasoning settings, none and high, a limitation noted by developer Simon Willison in his launch-day testing. Those settings give developers a choice between processing modes, but do not provide intermediate levels for tuning the balance between reasoning effort, latency and cost.

Mistral is emphasizing practical workflow performance rather than presenting the parameter count as proof of superior reasoning. The company claims a leading result on AutomationBench and lower per-task costs than competing models. Such claims need to be distinguished from general performance: benchmark results reflect particular tasks and evaluation conditions, while deployment costs also depend on prompt length, generated output and the number of attempts required to complete a job.

The forthcoming weights release remains separate from the API preview. The supplied reporting gives conflicting dates within late October, so an exact release day is not clear. Licensing terms also have not been finalized. Until those terms are published, the planned release should not be treated as a guarantee of unrestricted commercial use or described as open source solely because model weights will be downloadable.

For organizations considering self-hosting, access to weights would provide greater control over where inference runs and how data is handled. It would not eliminate infrastructure costs: operating a trillion-parameter model requires significant memory and serving capacity even when only a fraction of its parameters are active for each token. The license and deployment requirements will therefore be important details when Mistral makes the weights available.

Sources: Startup Fortune