Z.ai introduces GLM-Legal on GLM-5.3-Flash: 19.2% Harvey LAB all-pass (vendor-reported), API by application

Core Background & Industry Pain Points

LLMs are moving from code into broader professional work, and law is a high-stakes test: facts, statutes, reasoning, and a client-ready deliverable must link up and match the brief, while general models tend toward vague advice and unsound citations. On 2026-10-12 Z.ai introduced GLM-Legal (tech blog), a legal model built on the lightweight GLM-5.3-Flash.

Architecture Highlights & Internals

Training has two stages: mid-training on expanded legal corpora to strengthen fact-rule-conclusion reasoning, and post-training with RL across tens of thousands of realistic legal environments, co-evolving the model with its agent harness (context, tools, execution feedback) and turning practitioner reviews into training signals. Showcased tasks: contract review (tracked-change edits in Word), evidence analysis (mapping contradictions between external reports and internal records, with document IDs), and cross-border compliance planning (the blog's example covers Singapore, Malaysia, and Indonesia).

Authoritative Benchmarks & Measured Scores

All figures are vendor-reported, with no independent replication yet. Harvey LAB all-pass rate (24 practice areas): GLM-Legal 19.2%, Kimi K3 12.9%, Claude Fable 5 11.3%, GLM-5.3-Flash 6.7%, GPT-6 Astra 5.4%, a 12.5-point gain over the base. LegalBench and the Chinese PLawBench are only described as "parity" with frontier closed models, with no scores given. Even the top score means about four in five tasks don't fully pass, far from unsupervised delivery.

Developer Hands-on Guide

1. No open weights (no Hugging Face release); API access is by application, and pricing is undisclosed. 2. The desktop agent AutoClaw will get GLM-Legal "soon," with no date. 3. Evaluate blind on your own contracts and case files against lawyer annotations, checking citations closely. 4. Confirm data-transfer and confidentiality terms before using client data.