What Mistral released

On 29 October 2026, Mistral released Large 4 weights and published deployment options. It is a multimodal mixture-of-experts model with roughly 1.05 trillion total parameters and 52 billion active parameters. The release expands options for companies that need to host a model themselves, but downloadable weights do not mean it is practical to run the model on a single office server.

The vLLM instructions show an FP8 configuration using eight GPUs. An NVFP4 option is shown for four Blackwell GPUs; requirements and support depend on the format and hardware. These are published configurations, not universal minimum specifications for every workload.

Licence and infrastructure

The Mistral Ecosystem Open-Weight License permits use, modification, fine-tuning and commercialisation. Companies operating a Model-as-a-Service offering must obtain a separate commercial licence from Mistral. Before deployment, check the current licence text and the legal structure of your service.

Before testing the weights, assess GPU memory and accelerator compatibility, disk capacity, KV cache, inference-engine version and multimodal support. Include electricity, cooling, operations, support, redundancy and downtime costs.

For most SMEs, a sensible first step is an API or managed inference if data policy permits. If data cannot leave the company, compare a protected deployment with smaller local models on the same task set. A server purchase should be justified by actual workload, quality, latency and TCO—not merely by the availability of weights.

Test claims on your own tasks

Mistral’s benchmarks describe the developer’s evaluation, but do not show how Large 4 will perform on your internal documents. Build a task set from your own work and measure accuracy, cost per accepted result, latency and the share of answers corrected by employees.

Also assess safe updates, permissions, auditing, queues and request limits. An open-weight model still requires a mature operating environment and an accountable owner.

Practical takeaway

Large 4 is a significant release for organisations that need open weights and deployment control. For an SME, it is not an automatic reason to buy GPUs. First test an API or managed service, then confirm that local hosting is justified by data requirements, workload and total cost.