Z.ai launches GLM-Legal on GLM-5.3-Flash: 19.2% all-pass on Harvey LAB; API beta by application only, no open weights yet

Why it matters

【Key points】

Z.ai (Zhipu) on Oct 12 launched GLM-Legal, a legal-specialized model built on lightweight GLM-5.3-Flash. It reports a 19.2% all-pass rate on Harvey's Legal Agent Benchmark (LAB)—12.5 points above the base—across 24 practice areas, and claims parity with frontier closed models on LegalBench and Chinese-practice PLawBench at a much smaller parameter footprint.

Training mixes legal mid-training with RL across tens of thousands of legal environments, co-evolving the model with an agent harness for context, tools and delivery. Access today is API beta by application (z.ai/consultation/legal), with AutoClaw desktop demos for contract review, evidence analysis and cross-border privacy planning.

Caveat: Vals AI called it #1 open-weight on Harvey LAB, but the blog only offers API beta and there is no GLM-Legal repo under HF zai-org—do not describe this as an open-weight release.

【Verifiable】

Z.ai blog (dated 2026-10-12 · Research): https://z.ai/blog/glm-legal

@Zai_org post (2026-10-12 16:10Z): https://x.com/Zai_org/status/2109678174468534511

API apply: https://z.ai/consultation/legal

Harvey LAB 19.2% / +12.5pp and LegalBench/PLawBench parity are vendor claims; no independent replication yet. No HF weights found this patrol.

【Our read】

1. Another Chinese lab slicing a general flagship into a vertical agent: less about parameter count, more about domain mid-training plus harness co-training so the model ships usable work product, not just prose. Legal's fact–doctrine–argument–deliverable chain is a hard stress test.

2. Access is deliberately tight (apply-only API, no weights), which fits collecting production feedback and containing legal-hallucination risk. Repeating “#1 open-weight” from third parties would mislead.

3. The industry signal is professional cowork moving from demos toward billable workflows; watch API access, pricing and real firm uptake, not just LAB scores.