Open Source · OCTOBER 20, 2026
Mistral Large 4 'Le Chonk' preview goes live, open weights scheduled October 27
A 1.05T-parameter MoE with 49B active per token and a 1M-token context, in preview on Mistral Studio now. Weights ship by end of October — the test will be whether operators can reproduce the vendor benchmarks on their own hardware.
Mistral AI opened a public preview of Mistral Large 4 on October 6, with downloadable weights scheduled for release by the end of this month and a target self-hosting date of October 27. The company is calling the model "Le Chonk," which is a tell about who they think the audience is: not procurement officers at banks, but the open-weight developer community that has spent eighteen months watching DeepSeek set the pace from Hangzhou.
The spec sheet is the pitch. Large 4 is a 1.05-trillion-parameter mixture-of-experts with roughly 4.7% of weights activating per token, a 1M-token context window, multimodal input, and training across 160-plus languages. Preview API pricing lands at $1.36 per million input tokens and $4.18 per million output. Mistral says it trained the model on 3,800 NVIDIA Grace Blackwell GPUs over about two months.
The active-parameter count is where the launch materials start to wobble. BetaNews, working from current model documentation, lists 52 billion active parameters per task and notes that the 49-billion figure repeated by AI News and MarkTechPost appears to come from earlier materials no longer reflected on Mistral's official pages. Not a scandal. But a useful reminder that every number in an open-weight launch deck is the vendor's own until independent researchers get their hands on the checkpoint.
The benchmarks are the vendor's too. Mistral reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, and a combined 49.8% on the Coding Agent Index. On AutomationBench, which covers 657 business workflows across Gmail, Google Sheets, Slack, and Salesforce, Large 4 scores 59.9%. On Surge AI's blind human coding evaluation, it placed second of five models with a 3.74, behind Claude Opus 5 at 4.22.
The third-party number worth sitting with is the Vals Index: 48.05%, ranking 32nd of 44 models overall but 9th among open-weight entries. On Harvey's legal agent benchmark, Large 4 finishes first among open-weight models.
That's the shape of the claim. Not frontier-competitive on raw capability, but credibly top-of-class on the open side, at mid-tier API prices, with weights you can actually download. It's a positioning that echoes what Meta tried with Llama 2 in 2023 and what DeepSeek converted into genuine strategic alarm last winter: the argument that open weights plus serviceable benchmarks plus runnable inference economics is a product category of its own.
October 27 is when that argument stops being a slide deck. Operators will load the checkpoint, run their own evals, and the gap between 49 billion and 52 billion active parameters will either resolve itself or become the first footnote in a longer story.