Mistral AI opened a public preview of Mistral Large 4 on October 6, a sparse mixture-of-experts model it has nicknamed "Le Chonk," and committed to releasing downloadable weights by the end of October, with VentureBeat putting the specific drop date at the 27th. The headline claim is a 59.9% score on AutomationBench's 657-task suite grading frontier models on real business workflows, which, if independently reproduced, would make it the first Western-built open-weight model at trillion-parameter scale to compete with closed frontier systems on actual Gmail, Google Sheets, Slack, and Salesforce tasks.

The architecture is 1 trillion total parameters with 52 billion active per token in Mistral's own launch materials, though VentureBeat flagged a 49-billion figure from earlier documentation. Mistral trained the model on roughly 4,000 Nvidia Grace Blackwell GPUs across its European datacenters over about two months, per CNBC and VentureBeat.

Preview pricing on the hosted API lists at $1.36 per million input tokens and $4.18 per million output, with a launch-sale promotion cutting both in half to $0.68 and $2.09, per BetaNews.

Mistral's self-reported ranking places Large 4 ahead of Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro on AutomationBench. VentureBeat noted the model doesn't yet appear in Artificial Analysis's public evaluations, so the position is provisional until a third party reruns the suite. Context matters here: Gemini 4 Argon reset the top of that board earlier this month, so 59.9% is competitive with the closed frontier, not above it.

Pierre Stock, Mistral's VP Science, told TechCrunch the weights will ship after safety testing with trusted partners and governments. That phrasing is doing a lot of structural work. It gestures at the regulatory choreography European AI labs have had to perform since the 2024 AI Act implementation phase, where "open" and "government-reviewed" are no longer in tension by default.

For a five-to-thirty-person team weighing whether to run customer-facing classification, outreach drafting, or content automation on self-hosted infrastructure instead of paying per-token API rates, the arithmetic isn't yet available. A 1T-parameter MoE needs real inference hardware, and Mistral hasn't published a reference serving footprint or a cost-per-million-token figure for self-deployment. The halved preview API rate also changes the break-even math while it lasts.

The disciplined move is to wait three weeks. The weights land October 27, independent AutomationBench reproductions will follow within days, and GPU rental quotes for the active-parameter footprint can then be priced against a hosted rate that may or may not revert to $1.36. The open-weight story at frontier scale is finally real on this side of the Atlantic. The economics are still unpublished.

Sources

https://mistral.ai/news/mistral-large-4/
https://techcrunch.com/2026/10/06/mistrals-new-1t-model-aims-to-leapfrog-closed-and-open-rivals/
https://www.cnbc.com/2026/10/06/mistral-ai-model-le-chonk.html
https://venturebeat.com/technology/mistral-debuts-large-4-le-chonk-a-1-trillion-parameter-text-output-model-with-high-benchmarks-planned-for-open-weights-release
https://betanews.com/article/mistral-large-4-preview-open-weights/