Mistral releases Large 4 open weights for self-hosted multimodal workloads

The short version

Mistral has released Mistral Large 4 open weights. The 1T-parameter, 52B-active multimodal MoE can now be downloaded and self-hosted instead of used only through an API preview.

What changed

Mistral released improved Mistral Large 4 weights with an official Hugging Face download, model card, and deployment guidance. The model had previously been API-preview-only; the announcement also lists new hosted pricing and open-weight license conditions.

What it means for you

This is a delivered open-weight model release, not a routine feature update. Teams can now evaluate local or private-cloud deployment instead of only vendor API access, while still needing to review licensing, hardware cost, and production safety controls.

Mistral has released the weights for Mistral Large 4. This turns a previously API-preview-only trillion-parameter multimodal model into an open-weight option that teams can download, inspect, and deploy themselves.

Mistral describes this version as a 1T-parameter, 52B-active multimodal MoE for coding, reasoning, agent workloads, and defensive cybersecurity. Developers can obtain the weights from its official Hugging Face repository and run them on their own infrastructure with vLLM or SGLang. Use the current model card, actual hardware requirements, and a team's own tests as the source of truth for runtime compatibility.

Teams needing data control, private-cloud deployment, or customization can now compare self-hosting with Mistral's hosted API. The announcement lists hosted pricing of $0.70 per million input tokens, $0.07 cached input, and $2.10 output. The weights use the Mistral Ecosystem Open-Weight license, which the company describes as a modified MIT license; a model-as-a-service business needs a separate commercial license for commercial use.

Open weights do not remove production risk. Start in an isolated environment with a fixed task set, measure quality, latency, GPU memory, and total cost, and preserve least privilege, human review, and reversible execution boundaries for agents. Test prompt-injection and data-exfiltration defenses when workflows ingest untrusted code, tickets, or web pages.

Related directory entry

Fact check

VerifiedMistral released Mistral Large 4 weights on October 29, 2026 and links to an official Hugging Face download.Evidence

VerifiedMistral describes the released model as a 1T-parameter, 52B-active multimodal MoE and says it can be deployed on an organization's own infrastructure with vLLM or SGLang.Evidence

VerifiedMistral lists hosted prices of $0.70 per million input tokens, $0.07 cached input, and $2.10 output; a model-as-a-service business needs a separate commercial license for commercial use.Evidence