News · 2026-10-22

Mistral Large 4 is usable by API while its weights and license remain pending

Mistral Large 4 is available as a hosted public preview, but its downloadable weights were still pending on October 22. Mistral says they will arrive by month-end, while the eventual license and official self-hosting memory requirement remain unpublished in the checked materials. The distinction matters for customers evaluating European deployment control rather than simply another hosted model.

Key facts

• Mistral launched the public API preview on October 6, 2026.
• Its documentation lists 1.05 trillion total parameters and 52 billion active parameters.
• The Hugging Face page lists an upcoming release with an October 31 estimated arrival.
• The primary source is Mistral’s launch announcement.

Mistral’s phrase “by the end of the month” describes a forward commitment, not an artifact customers can inspect today. Its Hugging Face release page had no downloadable files at the dossier cutoff. There is consequently no verified weight-download size to report. The reviewed pages also give no minimum or recommended graphics-memory requirement, supported local runtime or measured self-hosted configuration.

The model, nicknamed Le Chonk, uses a mixture of experts. A routing mechanism chooses a subset of specialist networks for each token. Picture a large engineering company where a particular job goes to a few relevant specialists rather than every employee. That reduces work per job, but the company still needs access to the full staff for other jobs. Similarly, a small active subset does not make the full expert collection disappear from storage or the serving system.

The model documentation lists 52 billion active parameters and a 1.6-billion-parameter vision encoder. The release page separately describes 49 billion active parameters per token, or 52 billion including embeddings and output layers. Those figures use different accounting boundaries. Presenting them as an unexplained contradiction would obscure a distinction the primary material actually supplies.

Mistral describes a model spanning text, image understanding, reasoning and tool-driven workflows. It exposes two reasoning settings and highlights tasks such as reading engineering drawings, writing code and analyzing financial material. The company says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell accelerators in its European data centers. That is a company-reported training setup; it is neither an independently audited training account nor a specification for running the released model.

The performance evidence needs the same attribution. Mistral reports selected comparisons against other models and a blind coding-quality evaluation with Surge AI. These support reporting the company’s claimed strengths. They do not settle overall frontier leadership or establish reliability on a customer’s particular documents and tools. A long context window similarly describes capacity, not proof that every fact inside a large request will be used correctly.

Mistral’s live model page lists a one-million-token context window, while Artificial Analysis and Vals AI list 512,000 for the preview. The checked sources do not explain the discrepancy, so a usable one-million-token API request is not established. External evaluations also show task-dependent strengths rather than overall leadership; Artificial Analysis reports substantially higher task cost than the comparable open-weight alternatives it names. The model page lists prices of $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. Mistral’s changelog records a two-week half-price launch promotion. The October 22 rates are post-promotion prices, so old launch comparisons can mislead buyers even when copied accurately from an earlier announcement.

The sovereignty pitch is particularly relevant to security work. Mistral argues that organizations may need to reproduce vulnerabilities and test repairs without losing access through a hosted provider’s refusal or policy changes. If usable weights and suitable licensing arrive, self-deployment could give customers more control over those conditions. For now, that prospect depends on delivery, legal terms, runtime support and affordable infrastructure. Calling the preview unconditionally open source would get ahead of the available evidence.

Reception combines enthusiasm and doubt. Simon Willison welcomed Mistral’s return to stronger competition while placing it behind the frontier in his informal assessment. His single illustrative prompt showed little difference between reasoning modes; it cannot establish their general effectiveness. The launch discussion included support for European choice alongside questions about missing weights, licensing, hardware dependence and hosted value.

Mistral’s October 21 NVIDIA announcement gives the model a current industrial connection: engineering workflows combining Large 4 with simulation, design and accelerated-computing tools. It is not a new announcement of Large 4 training or serving hardware. The practical test remains concrete: verify the weight files, read the license and measure the intended workflow when the release lands. The open-weight distinction explains why those checks determine what customers actually control.

Primary source, verified: read the paper →

Ground Truth. (2026, October 22). Mistral Large 4 is usable by API while its weights and license remain pending. Ground Truth. https://groundtruth.day/news/mistral-large-4-preview-weights-still-pending.html

@misc{groundtruth:mistral-large-4-preview-weights-still-pending,
  title  = {Mistral Large 4 is usable by API while its weights and license remain pending},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/mistral-large-4-preview-weights-still-pending.html}
}

Topics: mistral · open-weights · model-release · enterprise · sovereignty