Z.ai Releases GLM-Legal, an Agentic Model Built for Legal Work
Built on GLM-5.3-Flash, GLM-Legal hits a 19.2% all-pass rate on Harvey LAB and matches frontier closed-source models on LegalBench and PLawBench.
Z.ai has released GLM-Legal, a domain model aimed at agentic legal work rather than single-turn question answering. It is built on the lightweight GLM-5.3-Flash base and trained in two stages: mid-training on expanded legal corpora, then reinforcement learning across tens of thousands of simulated legal environments.
The headline number is a 19.2% all-pass rate on Harvey LAB, covering real law-firm workflows across 24 practice areas — 12.5 points above its base model. Harvey’s grading is strict: a task counts only if every requirement passes, so the score is meant to show end-to-end reliability rather than partial credit. Z.ai also reports parity with leading closed-source frontier models on LegalBench and on PLawBench, a Chinese-language benchmark for consultation, case analysis, and drafting. Vals independently evaluated the Harvey LAB and LegalBench results.
The model is being wired into AutoClaw, Z.ai’s desktop agent platform, where published examples include contract review, evidence analysis, and pre-launch privacy compliance across Singapore, Malaysia, and Indonesia. API access is currently offered as a beta via consultation.