On this page
The short answer
Mistral Large 4 is a multimodal model available through Mistral’s API and as open weights under its stated license. The newer weight release adds reasoning-mode changes and reported coding gains. Evaluate the exact release and deployment you plan to use, with a fixed task budget and explicit acceptance rules.
What to take away
- Distinguish the public preview from the improved weight release when comparing results.
- Choose a reasoning mode deliberately and measure its effect on accepted tasks.
- Include serving and operating costs when comparing self-hosting with an API.
Sources checked . Research synthesis and Rubrex recommendations. Examples and local experiments are labeled; they are not client results.
A dated model snapshot. Published metrics describe the named model, settings, and test. They are not Rubrex experiments or a universal ranking. Prices and availability can change after the research date.
Compare models and check the data date →Identify the release before reading its score
Mistral’s open-weight announcement describes an improved Large 4 following its public preview. It reports 41% on Terminal-Bench 4.0, a gain of 13 percentage points over the preview. That comparison belongs to the provider’s release account; it is not evidence that every deployment or task improves by the same amount.
The announcement lists low, medium, and high reasoning modes, with high as the default and low replacing the earlier none mode. When you record an experiment, include this setting alongside the model identifier. Comparing modes at different time limits can answer a useful budget question, but it should not be presented as a pure model-quality comparison.
Evidence: Mistral Large 4 is now open weight [1]
Use a dated specification sheet
The model documentation checked on November 6 lists a one-million-token context window, 1.05 trillion total parameters, 52 billion active parameters, and a 1.6-billion-parameter vision encoder. It labels the release GA and names the Mistral Ecosystem Open Weight License v26.10. Review the actual license terms for the deployment you intend.
The same page lists API rates of $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output. Those are the provider’s displayed rates at review time. Keep them separate from the older OpenRouter snapshot on our comparison page; different retrieval dates and routes can produce different price records.
Evidence: Mistral Large 4 model documentation [2]
Compare deployment options using the same acceptance bar
Our proposed trial starts with the work the model must complete, rather than a preference for hosting. Choose representative inputs and a reviewer rubric. For a coding assistant, require a correct change, passing relevant tests, a readable explanation, and no unrelated edits. For document work, require each material answer to point back to the supplied evidence.
Run the hosted candidate and any self-hosted candidate against the same cases. Record the weight revision, precision, serving software, reasoning mode, and tool configuration. If the self-hosted version uses a different quantization or prompt template, make that difference visible. Matching a model name alone does not demonstrate that the two systems are equivalent.
Track operational work as well as usage charges: deployment time, capacity planning, monitoring, failed requests, and recovery. Put those costs beside the accepted-task count. A team with variable demand may reach a different decision from a team that can keep dedicated capacity busy.
Turn the trial into a repeatable decision
Include at least one case where the correct response is to stop or ask for missing information. A coding task can omit a required dependency; a document question can lack the requested fact. Check whether the agent reports the limitation clearly and preserves the working state, rather than claiming completion.
Review failures by consequence. An incorrect explanation, an unauthorized action, and an unresponsive deployment need different responses. Agree who decides whether a failure blocks release and what evidence closes it. Keep these decisions with the task record so the next model or configuration change can be assessed against the same standard.
Choose the deployment only after comparing accepted outputs, completion time, operating effort, and the most serious failures. Keep a small regression set for subsequent upgrades. The purpose of this process is to make the tradeoff inspectable, not to convert one public benchmark into a universal ranking.
Limits of the evidence
Specifications and release claims were checked against Mistral’s official pages on November 6, 2026. This is a source-based briefing, not a Rubrex benchmark or a license opinion. Hosting requirements, price changes, and suitability for a particular workflow need separate verification.
Common questions
Does open weight mean self-hosting is free?
No. Compare infrastructure, operations, and review effort alongside license terms. The decision should use your expected workload and accepted outcomes.
Should every task use high reasoning?
Test the available modes against your acceptance criteria and latency budget. The default is a starting configuration, not evidence that it is best for every task.
Sources & further reading
Primary sources behind this briefing. A source’s findings apply to its own study conditions; publication on arXiv does not establish peer review.
- Mistral Large 4 is now open weight Mistral AI · 2026 · Release announcementReviewed: updated weight release, reasoning modes, and attributed Terminal-Bench result. Accessed November 6, 2026.
- Mistral Large 4 model documentation Mistral AI · 2026 · Maintained documentationReviewed: capacity, license label, parameter counts, and dated first-party API rates. Accessed November 6, 2026.
Questions or corrections? Write to Rubrex. Read our editorial approach.