Mistral releases Large 4 weights for self-hosted deployment

Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on October 29. Developers and organisations can now download, modify and deploy the model on their own infrastructure. The company says additional reinforcement learning since its public preview has improved coding and defensive cybersecurity capabilities.

The release follows the API preview announced on October 6. Mistral initially offered online access to its roughly one-trillion-parameter model and promised the weights by the end of the month. Those parameters are the numerical values learned during training. Distributing the files gives organisations a way to operate the model independently of the original provider’s hosted API. The preview announcement listed API rates of $1.36 for input and $4.18 for output per million tokens.

The official model card specifies text and image input, text output and a context window of roughly one million tokens. It distinguishes 49 billion active parameters per token from 52 billion when embeddings and output layers are included. These are two accounting scopes for the same model.

Requests can select low, medium or high reasoning effort. It also supports function calling, structured JSON output and, according to Mistral, more than 160 languages including all official EU languages. Image capabilities cover documents, charts and technical drawings. Function calling and structured output support integration with other software.

Hugging Face’s deployment guide identifies two official weight formats, FP8 and NVFP4, which use different numerical precision to adjust storage and inference requirements. Community-produced GGUF conversions are also available. Alongside downloadable files, the guide describes deployment through hosted inference endpoints, offering separate routes for organisations that want to operate the model themselves or use a managed service. The NVFP4 release targets NVIDIA Blackwell GPUs, with separate guidance for that format and GGUF-based operation.

Memory and licensing for self-hosting

Running the model requires substantial memory. The vLLM recipe lists approximately 1.06 terabytes of FP8 weights and 610 gigabytes for NVFP4, before additional runtime needs such as the memory used to retain conversation context. It documents validated server-GPU configurations. Although the model advertises a million-token context, this recipe says it validated up to 256K tokens, a narrower tested range. Validated hardware includes NVIDIA GB300 and B200 systems and AMD MI350X accelerators; configuration depends on precision and context requirements.

The Mistral Ecosystem Open-Weight License permits use, modification, further training and distribution, with copyright and permission notices retained. Internal use that does not make the model or derivatives available to third parties is exempt from its separate commercial-licensing requirement.

A licensee or affiliate operating a Model-as-a-Service business needs a separate agreement before commercial use. This covers third-party inference or post-training with meaningful control over inputs, parameters or training data. It excludes capabilities embedded solely within specific product features and merely relaying requests to models hosted by others.

For customers using the hosted API, the current product documentation lists prices per million tokens of $0.70 for input, $0.07 for cached input and $2.10 for output. It also lists document question answering, batch processing and agent-conversation interfaces. These API charges are separate from the infrastructure costs of a self-hosted deployment.

A separate cybersecurity variant is in preview with enterprise design partners. Mistral requires vetting and contractual safeguards, restricting tests to owned or expressly authorised systems in controlled environments.