Mistral Large 4 expands self-hosted AI options—but licensing matters

What happened

Mistral released the downloadable weights for an improved version of Mistral Large 4, also known as “Le Chonk,” on October 29. The multimodal model is designed for coding and agent workflows and is available both for self-hosted deployment and through Mistral’s hosted API. Mistral describes the release as an open-weight model, but that does not mean unrestricted open-source use.

The model card lists approximately 1 trillion total parameters, with about 49 billion active per token—or 52 billion including embedding and output layers—and a 1 million-token context window. Those specifications point to serious infrastructure requirements. This is not a model that should be assumed to run on an ordinary business laptop or a modest server.

The accompanying license permits internal use and adaptation, but requires a separate commercial agreement for certain Model-as-a-Service businesses. The restriction applies when a licensee or its affiliates commercially operate such a service using the model. Internal use, official Mistral products, embedded end-user features and simply relaying requests to externally hosted models are excluded from that definition.

Mistral’s announcement highlights improvements in coding, cybersecurity and safety. Those claims are vendor-reported; the release information does not establish independent comparative superiority.

Why it matters

For companies with genuine data-control, customization or deployment requirements, Large 4 adds another substantial option beyond a managed API. Self-hosting can reduce dependence on a provider’s inference endpoint and give technical teams more control over where sensitive workloads run. It can also create a new operating burden: hardware, deployment, monitoring, upgrades and engineering time become the customer’s responsibility.

The economics therefore depend on workload fit, not model size alone. A business running occasional document analysis may find a hosted API cheaper and simpler. A company with steady volume, strict data-location requirements or a product that embeds model capabilities may have a stronger case—but must still validate performance on its own tasks.

The license is part of that evaluation. A software company planning to offer model access as a service should obtain legal review before building around the weights. Mistral’s release coverage confirms the significance of the launch, but the practical decision remains specific to each business’s workload and risk profile.