Z.ai Launches GLM-Legal Legal Agent With 19.2% Harvey LAB All-Pass Rate

What happened

According to Z.ai's official blog announcement on October 12, GLM-Legal is an agent model built for legal workflows: extracting core issues from source documents, building defensible arguments, and formatting deliverables to task specifications. The company describes legal work as an ideal stress test for training models to operate in high-stakes professional environments, where facts, statutory rules, analytical reasoning, and client-ready work product must form an unbroken chain.

The model is built on a 320-billion-parameter mixture-of-experts architecture (18 billion active per token) with a 1-million-token context window. Training ran in three stages: mid-training on expanded high-quality legal corpora, post-training reinforcement learning across tens of thousands of realistic legal environments, and converting production feedback from legal professionals into training signals.

Key facts

1. Harvey LAB: 24 practice areas under strict all-pass grading, 19.2% pass rate, 12.5 points above the GLM-5.3-Flash base (company figures, no third-party replication yet).
2. LegalBench / PLawBench: the company claims precision on specialized contracts, merger filings, and judicial opinions on par with top closed-source frontier models, and parity with leading frontier systems on PLawBench (Chinese legal consultation, case analysis, drafting); comparison systems and scores were not disclosed.
3. Access: no public API yet; law firms, legal-tech companies, in-house legal teams, and independent practitioners can apply via a consultation form; the model will also ship inside the AutoClaw desktop agent (Windows / macOS / iOS / Android, 50+ built-in skills).