Mistral Large 4

Published frontier peer. Public preview nicknamed Le Chonk. Preview API only, in Mistral Studio. Weights promised by the end of this month (announced open-weight); not downloadable yet. No license is stated. 1T total / 52B active parameters (architecture details not yet published). Natively multimodal. Vendor-claimed cites, not independently verified. DeepSWE v1.1 61.7% is the banked depth row and not an Index column. Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. SWE-Atlas-QnA 59.4% is a note only (no catalog bench). Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. Terminal-Bench 4.0 28.3% is a note only (chart label Terminal-Bench 4.0; not Terminal-Bench 2.1 and not an Index cell). Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. AutomationBench 59.9% is a note only. The post describes 657 business workflows and names no split, version, or protocol. Protocol differs from our desk runs. Not labeled the public split, and not filed on the local AutomationBench bench, so it is not a ranked row there. Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. Mistral's own Coding Agent Index 49.8% is a note only. Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. Cybench 93% (40 exercises) is a note only. Vendor-claimed. Harness, reasoning effort, sampling and sample count not stated. Index cells for SWE-bench Verified, Terminal-Bench 2.1, GPQA, and IFBench stay empty. Local fit is pending weights and untested. By size alone, 1T total parameters will not fit one GB10 (128 GB) or two GB10s (256 GB) even at low bit. No weights or community quants exist yet, so there is no EXAMPLE desk row. Not a LocalsOnly desk run. https://mistral.ai/news/mistral-large-4/