The downloadable checkpoint changes the deployment choice
Mistral has posted the Large 4 model weights for download after introducing the model through its API earlier in October. The issuer's model card describes a multimodal mixture-of-experts system with about one trillion total parameters, roughly 52 billion active parameters and a one-million-token context window. It accepts text and images and can return text, structured output and function calls.
The repository documents deployments through vLLM, SGLang, Transformers and llama.cpp. It also provides a reduced-precision checkpoint for NVIDIA Blackwell hardware. Those instructions make self-hosting possible, but they do not make the model lightweight: operators still need to size accelerator memory, storage, networking and serving capacity for their workload. CGH has not downloaded or benchmarked the weights.
The license is broad, with a hosting exception
Mistral distributes the checkpoint under its Ecosystem Open-Weight License, a modified MIT license. The text permits commercial and noncommercial use, modification and redistribution, subject to notices. It separately requires a commercial agreement when a business offers the model or a derivative as a service that gives third parties meaningful control over inference or post-training.
The license excludes internal use and some end-user products with embedded model features from that model-as-a-service condition. That distinction can affect hosted API, platform and reseller designs. It is a contractual term, not a technical restriction, and organizations should review the full license for their planned distribution model rather than infer permission from the availability of the files.
The hosted API has different operating constraints
Mistral's API documentation still labels Large 4 a public preview and lists usage-based token pricing. Optional regional inference is documented for the European Union and United States at a premium over the global endpoint, but model availability varies and stateful features are not offered on those regional endpoints. Mistral also says some account, billing and analytics information can remain outside the selected inference region.
The public model repository was available when CGH checked it on October 29, but its page does not supply an exact release time. Mistral's changelog still lists the October 6 API preview as its latest Large 4 entry. CGH has not verified the issuer's performance, cybersecurity or efficiency claims and has not assessed the model's safeguards.