Mistral releases Large 4 weights for self-hosted AI

Mistral has released the model weights for Large 4, nicknamed Le Chonk, following its public preview. Its announcement gives organizations a choice: use hosted access through Studio or the API, or download the weights from Hugging Face and deploy them with vLLM or SGLang. The latter are software options for serving the model on an organization’s own infrastructure.

The model card describes text and image inputs, with text responses, plus function calling and structured JSON output. Mistral recommends high reasoning effort for complex prompts and agent workflows, low for quick chat, and medium for everyday use; high is the default. For applications that call tools, its advice is to describe the environment, task and tool usage in the system prompt, and expose only the minimum necessary set of clearly defined tools.

Self-hosting has concrete deployment conditions. The current vLLM recipe specifies a dedicated Large 4 container image and a separate ROCm image for AMD; it says ordinary release and nightly wheels do not yet include support. Its validation covers specified NVIDIA server configurations and AMD MI350X hardware. Although the checkpoint advertises roughly one million context positions, the recipe reports validation only up to 256k and sets its maximum accordingly. That distinction matters when choosing how much material to submit in one request. AMD MI300 hardware remains unvalidated in this recipe.

The Decoder adds reporting from a press call: Mistral researchers Pierre Stock and Corentin Petit described internally created training tasks that became progressively harder. They linked the process’s limit to the difficulty of tasks the team could still create and verify. That account describes the team’s training approach, rather than an independent performance test.