Mistral Released Mistral Large 4's Open Weights - the Token-Efficiency Number Is the Real Story

On October 29, 2026, Mistral AI published the open weights for Mistral Large 4, the 1-trillion-parameter natively multimodal mixture-of-experts model it previewed earlier in October. The open-weight frontier has a new leader, and the interesting part is not the parameter count - it is what Mistral says it does with fewer tokens per coding task.

What Actually Shipped

The weights release is the news, not the size of the model. Mistral previewed Mistral Large 4 publicly on October 6 and said weights would follow by the end of October. They arrived on October 29, and Mistral says the roughly three weeks in between went into additional reinforcement learning.

That delta is measurable. Mistral says the released weights improved Terminal-Bench 4.0 by 13 percentage points over its own public preview. A version number matters here: the preview and the released weights are different models on every benchmark that follows, and quoting a score without saying which one you mean is how a 13-point improvement gets mistaken for a rounding error.

The Coding Numbers, Each Attributed

Every benchmark figure below is vendor-reported by Mistral in its own announcement. None of it has been independently reproduced.

FrontierCode 2.0 - 19.7%. Mistral says Mistral Large 4 leads this benchmark. It is Cognition’s private benchmark, so nobody outside Cognition can re-run it. Read the number as a claim from the vendor, not as a measurement you can check.
Artificial Analysis Coding Agent Index - 48%. Mistral reports this score as effectively tied with GLM-5.3, the leading open-weight model, while using 47% fewer tokens per task on average. The token figure is the load-bearing number in this post.
Terminal-Bench 4.0 - 41%. Mistral reports the released weights reach 41%, a 13-percentage-point gain over the preview, take the lead among open-weight models, and do it while cutting output tokens by 30%.

So the picture is a model that is not dramatically better than its nearest open-weight competitor on aggregate score, and is dramatically cheaper in tokens. That is a different and arguably more useful claim than “best model.”

Three Reasoning Modes Are the Feature You Can Actually Use

The shipped model carries three reasoning modes: Low for simple tasks, Medium for everyday work, and High by default for the hardest cases.

This is the section a platform team can act on today. Most of the cost of an agentic coding loop is the reasoning pass, not the answer, so a mode switch that lets a team send routine refactors through Low while reserving High for genuinely hard diffs is a direct lever on spend. Mistral does not publish per-mode pricing in the announcement, so the honest framing is that the switch exists and the cost difference has to be measured on your own traffic rather than read off a table.

What Sovereign AI Means Here, in Practice

Mistral says Mistral Large 4 was trained from scratch on European infrastructure, and it positions the release as a foundation for “sovereign AI” - downloadable weights, self-hosted deployment, and customization that does not depend on an API provider’s terms, quotas, or pricing.

That is a concrete contrast with API-only access to a comparable model. A downloadable checkpoint can be quantized, run on infrastructure you control, and kept as a fallback when a hosted provider moves its price or throttles your account. Mistral’s framing is its own, and the sovereign framing should be read as vendor positioning rather than as a policy outcome - but the self-hosting option is real regardless of which label you attach to it.

The Defensive Security Angle, Carefully

Mistral’s announcement frames the model for defensive security work, which follows from its earlier preview positioning. Secondary reporting from shattered.io gives the release a 51% score on the Artificial Analysis Cyber Index and a 96.6% refusal rate on malicious cyber requests. Those two figures come from that outlet rather than from Mistral, and they should be treated as unverified.

Note also that a 51% Cyber Index figure is not comparable to the 82% Mistral cited for the preview in early October - different numbers from different sources for the same benchmark family, on different model versions. If you are tracking this model, keep the versions separate.

What Open Weights Does Not Mean

Three caveats worth stating plainly.

First, open weights is not open source. The license terms for Mistral Large 4 were not verified in this research pass, so do not assume OSI approval, unrestricted commercial use, or redistribution rights.

Second, Mistral does not disclose the active-parameter count in the open-weight announcement. Secondary sources put it near 49-52B active against a total of roughly 1.05T; the official announcement simply says “1-trillion-parameter.” Treat any active-parameter figure as a secondary-source number.

Third, do not infer inference cost from the 1T headline. A trillion total parameters says very little about what a token costs to produce. The token-efficiency figures Mistral reports - 47% fewer tokens per task, 30% fewer output tokens - are the figures that bear on cost.

Why This Matters to a Platform Team

Three practical takeaways. First, the reasoning-mode switch gives you a cost control that does not require changing models or renegotiating an API contract. Second, a downloadable checkpoint is a real hedge: a fallback path that exists on your own hardware rather than on someone else’s rate limit. Third, Mistral separately announced a EUR 3B Series D at a post-money valuation above EUR 21B on the same day, which is the funding behind European training infrastructure and the reason the sovereignty claim has substance behind it.

The Number to Put in Your Own Eval Harness

The open-weight frontier moved on October 29. Of the three Mistral numbers, the one worth putting into your own evaluation harness is the token-efficiency figure, because that is the one you can measure directly against your own workload - and it is the one that decides whether a self-hosted agent is cheaper than an API call at your volume.