What happened
On October 29, Mistral announced the release of the Mistral Large 4 (Le Chonk) weights. Until then, the model had been available as a public preview through the company’s services; developers can now download the official checkpoint and deploy it themselves. The Hugging Face card provides the model files and serving instructions.
Large 4 is a multimodal Mixture-of-Experts model with about 1 trillion parameters in total. It activates 49 billion parameters per token (52 billion including embeddings and output layers). The model card lists 128 routed experts, four active experts per token plus one shared expert, a context window of up to one million tokens, image input, tool calls, and strict JSON Schema output. vLLM, SGLang, Transformers, and llama.cpp are supported; the NVFP4 checkpoint specifically targets NVIDIA Blackwell accelerators.
What this changes for businesses
Organizations that need control over model placement, connections to internal systems, and data processing now have another option to evaluate. Potential tasks include helping developers work with internal code, finding vulnerabilities in company-owned repositories, analyzing long technical documents, and running agent workflows with a limited set of tools.
Open weights do not mean that the model is easy or inexpensive to run on an ordinary server. Mistral’s vLLM instructions recommend a tensor-parallel configuration across eight GPUs and reserve 15% of device memory; this is an example configuration, not a universal minimum requirement. For SMEs, this is closer to a data-center than an office deployment. Most SMEs should first consider an API or managed inference and assess self-hosting separately, including hardware, electricity, operations, and utilization.
License and limitations
Mistral describes the Mistral Ecosystem Open-Weight License as a modified MIT license and permits commercial and non-commercial use. However, operating the model as a Model-as-a-Service business requires a separate commercial license from Mistral. Before deployment, check the license text against the intended distribution and operating model.
Independent Artificial Analysis lists a score of 26 for Large 4 on its Intelligence Index v5. Coding, cybersecurity, and efficiency figures in Mistral’s announcement are developer-reported claims and should be distinguished from independent evaluation.
Practical next step
If an on-premises environment matters, start with a small, closed set of tasks and compare Large 4 with an already available model on accuracy, latency, cost per successful result, and GPU requirements. For cybersecurity, test only systems you own or are explicitly authorized to assess; model responses should not automatically change code or access rights. If an API can answer the initial question, calculate the cost of local infrastructure only after the business value has been demonstrated.