Z.ai announces GLM-Legal legal agent on GLM-5.3-Flash

Z.ai's GLM-Legal posts 19.2% on Harvey LAB; a later headline claims wins over rivals without supporting scores.

Z.ai announced GLM-Legal, a domain model on the lightweight GLM-5.3-Flash base for agentic legal work such as contract review, case preparation, and document drafting. On Harvey LAB it recorded a 19.2% all-pass rate across 24 practice areas, 12.5 points above its base model, and the company says it matches frontier closed models on LegalBench and Chinese PLawBench with a leaner parameter budget. Training pairs expanded legal mid-training with reinforcement learning over tens of thousands of legal environments, co-evolved with an agent harness, and the model is slated for AutoClaw. A later MarkTechPost headline claims GLM-Legal beat Claude Fable 5 and Kimi K3 on Harvey LAB, but that page body was only a cookie-consent notice, so those named comparisons, scores, and model sizes are not confirmed in the retrieved text and conflict with the more specific HN figures that do not name those rivals.

• Z.ai announced GLM-Legal, a legal agent built on the lightweight GLM-5.3-Flash base, for contract review, case preparation, and document drafting (HN, 2026-10-12).
• On Harvey LAB it scored a 19.2% all-pass rate across 24 practice areas, 12.5 points above its base model.
• Z.ai says it matches frontier closed models on LegalBench and Chinese PLawBench with a leaner parameter budget.
• Training combines expanded legal mid-training with reinforcement learning over tens of thousands of legal environments, co-evolved with an agent harness.
• The model is planned for AutoClaw, Z.ai's desktop agent platform.
• A 2026-10-13 MarkTechPost headline says GLM-Legal beat Claude Fable 5 and Kimi K3 on Harvey LAB, but the retrieved page was only a cookie notice and contained no scores, sizes, or test details.