Mistral Opens Large 4 Preview, a Trillion-Parameter Model With Open Weights Due in October

Mistral AI has opened a public preview of Large 4, a multimodal model of roughly one trillion parameters that it describes as open-weight. The company announced it on October 6 and says it will release the weights by the end of the month (VentureBeat).

<blockquote><strong>Update, October 25:</strong> As of reports dated October 23, the weights and the licence had still not been published. VentureBeat reported a release date of October 27, while Mistral's own announcement commits only to the end of October. Mistral's documentation lists Large 4 as "Open" without naming a licence, and reports point to a custom Mistral licence rather than the Apache 2.0 licence used for Large 3 (<a href="https://forkast.news/the-license-countdown-mistral-large-4-and-the-20m-question/">Forkast</a>, <a href="https://letsdatascience.com/blog/mistral-trillion-parameter-weights-this-month-license-blank">Let's Data Science</a>).</blockquote>

What Mistral says it built

Mistral describes Large 4 as a hybrid instruct-and-reasoning mixture-of-experts model that takes text and images as input. The company puts it at about one trillion parameters with 52 billion active at a time, and says it was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs. It supports more than 160 languages, including every official language of the European Union.

Outside coverage uses slightly different figures. Let's Data Science, citing Mistral's documentation, gives 1.05 trillion parameters with 49 billion active, and VKTR also reports 49 billion active. Mistral's own materials appear inconsistent, so the precise count is best treated as unsettled. Let's Data Science also reports a context window of one million tokens, but Mistral's announcement page does not state a context length, so that figure is unconfirmed.

Price and access

The preview runs through Mistral Studio, with documentation at docs.mistral.ai, and is served from European data centres. Mistral says a multi-region rollout is planned, including a deployment operated by Mistral itself in Europe.

Model	Input, per million tokens	Output, per million tokens
Mistral Large 4 (preview)	$1.36	$4.18
Google Gemini 3.6 Flash (per VKTR)	$1.50	$7.50

The comparison covers list prices only, and the two models are not equivalent in size or design.

Mistral's benchmark claims

Test	Large 4 result (Mistral)	Notes
DeepSWE v1.1	61.7%	Software engineering
Coding Agent Index	49.8%	Mistral's own index
Cybench	93% of 40 challenges	Security
Artificial Analysis Cyber Index, reproduce-and-patch	82%	Claude Opus 5.5 and GPT-6 Astra refused this test, per VKTR
Dense 200 visual grounding	42%	GPT-6 Astra: 41%
Blind human coding evaluation	3.74, 2nd of 5	Claude Opus 5: 4.22, per VKTR

The refusals on the cyber test reflect those models' safeguards as much as their ability, and a one-point gap on visual grounding is too small to separate the two models. Nearly all of these figures come from Mistral, and its page does not describe the testing conditions. VKTR notes that a third-party evaluator found Large 4 ahead of GPT-6 Astra on legal and finance tasks but gave no numbers. Mistral adds that its reinforcement learning run is still going, so results may change before the final release.

What is still missing

The most important unanswered question is the licence. Mistral calls the model open-weight but has not named the terms. Open weights can mean anything from a permissive licence that allows commercial use to a restricted one with usage limits, and the difference decides who can build on the model. Mistral has said it will publish architecture details and more benchmarks alongside the weights.

There is also a difference between a preview and a release. Let's Data Science points out that hosted preview access and downloadable weights are separate stages, and only the first exists so far. Until the weights appear, nobody outside Mistral can run Large 4 on their own hardware or check its behaviour independently (Online Tech Tips).

Why Europe is watching

Mistral has built its pitch around European data residency and sovereign AI, and Large 4 leans on that. Running a trillion-parameter model is expensive, so the active-parameter design matters: only a fraction of the model works on any given token, which keeps serving costs closer to those of a much smaller system. Whether the quality holds up outside Mistral's own tests is the open question.

What to watch next

Watch for the weights and their licence by October 31, and for independent benchmark runs once developers can test the model themselves. Watch whether the parameter and context figures are corrected in Mistral's documentation, and whether the pricing holds once the preview ends.

Related on Gadgets365

• Meta releases Muse Spark Open, an 816-billion-parameter open-weight model
• Meta releases Muse Wonder and prices the API at $4 per million input tokens
• Google unveils Gemini 4 Argon