Mistral Opens Access to Large 4, but Keeps the Weights to Itself Until Red-Teaming Finishes

Mistral AI has opened a guarded public API preview for Mistral Large 4 — a natively multimodal mixture-of-experts model the company describes as having one trillion total parameters, with 49 billion active per inference. What you can't do yet is download it: the weights aren't public, so no developer can currently run the model independently.

That distinction matters more than it might sound. This is an API launch, not an open-weight release. Mistral controls the endpoint, the access terms, and the moderation around it — none of which is the same as publishing a model package that outside organizations can inspect and operate under their own terms. Mistral says open weights are coming by the end of October, after red-teaming. TechCrunch reports roughly three weeks post-preview following safety testing; Axios reports a more specific October 27 target, tied to further reinforcement-learning and safety work. Those two timelines roughly agree, but neither is a guarantee — both describe a plan, not a completed release.

Worth being clear about what's actually confirmed here versus what's missing. The 1T total / 49B active parameter split is Mistral's own stated architecture — a real, specific figure, not vague marketing language. But the available reporting doesn't include licensing terms, pricing, hardware requirements, context length, training data, or any benchmark results. There's also no independent verification of how Large 4 performs against competing models — any performance comparisons so far are Mistral's own framing.

Bottom line: Mistral has made a concrete architectural claim and opened real API access to back it up — that's more substantive than a bare announcement. But the actual test of this release is still ahead: whether the open weights land on schedule in late October, and once they do, whether independent benchmarks and real-world use confirm what the sheer parameter count implies about capability.