Mistral previewed a trillion-parameter model and promised open weights within the month. It has not said yet what licence those weights will carry.

Mistral AI put Large 4 into public preview on 6 October 2026, reachable now through the Mistral Studio API console. Mistral’s own announcement describes it as “a 1 trillion-parameter natively multimodal model with 52 billion active parameters,” trained on 3,800 Nvidia Grace Blackwell GPUs in European data centres and covering more than 160 languages, including every official European Union language.

Preview API pricing is set at 1.36 dollars per million input tokens and 4.18 dollars per million output tokens. Mistral says open weights are coming “by the end of the month” — its own wording, which commits to end-October 2026 without a firmer date, though at least one outlet has reported a specific 27 October release that Mistral’s own announcement does not state.

Why it matters

A trillion-parameter model trained entirely on European infrastructure, and pitched explicitly around European data-sovereignty, is a different buying proposition than the same capability from a US or Chinese lab, for any buyer where data residency or export-control exposure is a procurement requirement.

The licence question matters because it decides who can actually use the open weights commercially once they ship, and Mistral has a history of varying its terms by model rather than defaulting to one licence family, which means the Large 3 precedent (Apache 2.0, unrestricted) cannot simply be assumed to carry forward.

The licence is the open question, and Mistral has not answered it

Mistral’s own developer documentation lists Large 4 with an openness status of “Open” but names no specific licence, and the announcement page itself only commits to releasing weights by month’s end without licence terms attached. That is a meaningfully different position from Mistral Large 3, which shipped under the permissive Apache 2.0 licence, and from Mistral Medium 3.5, which 18BYTE’s own profile flagged as shipping under a modified MIT licence that is “not plain MIT and needs reading.”

Until Mistral publishes the actual licence text alongside the weights, nobody outside the company can say whether unrestricted commercial self-hosting will be allowed, or whether the terms will carry the kind of field-of-use or scale restrictions that have appeared in other labs’ “open-weight but not open-source” releases this year.

Benchmark claims are Mistral’s own, and specs disagree across sources

Mistral reports Large 4 scoring 61.7 percent on DeepSWE v1.1, a software-engineering task-completion benchmark, 28.3 percent on Terminal-Bench 4.0, which measures multi-step terminal workflows, and 93 percent on Cybench, a set of security-competition exercises. It also claims a top-five global placement on Artificial Analysis’s Cyber Index, a vulnerability detection and patching benchmark that is Artificial Analysis’s own index rather than Mistral’s. All of these are vendor-reported, with no independent benchmark run found to confirm them.

Mistral’s own announcement does not state a context window figure anywhere in its text. Secondary sources disagree sharply on this point: several report a 1 million-token context window attributed to “Mistral’s own model page,” while OpenRouter’s model registry listed a 524,288-token (512K) limit as of this preview. We could not resolve this discrepancy against Mistral’s primary announcement, so context window is left unstated here rather than guessed. Several outlets have also circulated a rounded 1.05 trillion-parameter, 49 billion-active-parameter figure for Large 4; Mistral’s own announcement states 1 trillion total and 52 billion active, and that is the figure used throughout this piece.

What this means for buyers

Buyers with a data-sovereignty requirement should treat this as a credible European-infrastructure option worth evaluating in preview now, but should not commit to self-hosting plans until Mistral publishes the actual open-weights licence at the end of October. Anyone citing Large 4’s parameter count or context window should go to Mistral’s own announcement rather than an aggregator, given the discrepancies already circulating on both figures within forty-eight hours of launch.

What would change our reading

Publication of the actual licence text when weights ship, expected by the end of October 2026, is the single fact that resolves the open question in this piece; we will revisit this story then. Mistral confirming a specific context window figure, and clarifying why OpenRouter’s registry and the figures attributed to Mistral’s own model page disagree, would also settle a loose end this draft could not close.