An API preview lets you examine model behavior. Planning a private deployment needs a release package you can inspect and run.
What is available now
Mistral announced Mistral Large 4 on October 6, 2026, with a public preview through Mistral Studio. The company plans to release the weights by the end of the month. That is a future release commitment, not a verified download available today.
The model documentation lists a 1-million-token context window and a mixture-of-experts architecture. It describes 1.05 trillion total parameters, with 52 billion active parameters. The preview supports multimodal work.
Keep preview costs separate from hosting costs.
On October 10, the documentation displays sale rates of $0.68 per million input tokens, $0.07 for cached input, and $2.09 for output. Confirm the rate in your account before budgeting. The displayed promotion does not establish a permanent price.
An API invoice and a private deployment have different costs. For self-hosting, a team also needs hardware capacity and a supported inference stack. The active parameter count alone cannot establish memory requirements.
Prepare a task set while waiting for artifacts.
For document work, choose questions with answers your team has already verified. Include scanned pages and cases where the answer is absent. Record whether the model cites the right evidence or invents missing details.
For coding, use a fixed repository snapshot. Review every proposed change and run the project tests. Record failures as well as successful tasks. This proposed trial does not reproduce Mistral’s published benchmarks.
- Record the preview version and request settings.
- Define acceptance rules before comparing models.
- Measure retries, latency, and accepted-task cost.
- Inspect the eventual weights, license, and model card.
- Test a supported serving configuration before reserving production capacity.
Company evaluations need a workload test.
Mistral publishes benchmark claims and says training behind the preview is still underway. Treat those results as company reporting. A later checkpoint can differ from the preview you tested; retain the version and date with your results.
Our Beam release checklist explains artifact checks. The model-routing guide covers workload-specific comparisons.
Wait for a reproducible deployment.
I would prepare the evaluation now and commit to hosting after reviewing the release package. Which document or repository task would expose a failure that your team cannot accept?
