Mistral Large 4 Weights Are Here. Check the License Before Hosting.

2026-10-30

Maestro’s take. Mistral Large 4 is worth testing for coding and defensive security work. The useful change is control over deployment, alongside a cheaper API. Start there. A downloadable model does not make serving it cheap or its license unrestricted.

TL;DR

Mistral’s October 29 release makes the improved Large 4 weights available. It lists API prices per million tokens of $0.70 input, $0.07 cached input and $2.10 output.

The model card describes a one-trillion-parameter model and documents self-hosting with vLLM and SGLang. This is infrastructure to size and operate, not a promise of laptop inference.

Read the Mistral Ecosystem Open-Weight License. If you or an affiliate operates a Model-as-a-Service business, it requires a separate commercial license before commercial use. Its definition excludes capabilities embedded solely within specific end-user features or harnesses, and mere relaying to someone else’s hosted models. It also exempts internal use and use through Mistral’s official products.

Why Maestro users care

Owning a deployable checkpoint gives a team another option when provider access, data location or customization matters. It also hands the team the serving bill and operating work. Those belong in the comparison.

Mistral reports coding and cyber benchmark gains. Treat those as vendor evidence, not a forecast for your repository. Nor do the public weights imply access to its separately vetted, specialized cyber variant, which the announcement describes as a partner preview.

Maestro would compare the hosted API first. Move to self-hosting when a concrete control requirement or measured workload justifies the extra work. For a model-service business, resolve the license question before building the offer around these weights.

One thing to try

Run a small set of your own coding or authorized vulnerability-fixing tasks against the new API. Keep the acceptance tests fixed. Record cost per accepted change, retries and review time. Then price a self-hosted pilot with enough capacity for your actual concurrency.

Published October 30, 2026 · Brief and Maestro commentary · Tags: models, costs, security

Original sources: Release and API prices · Model card and serving examples · Full license