GLM-Legal
Points and comments are a snapshot, not live.
GLM-Legal matches frontier models on legal benchmarks with a smaller model footprint.
Zhipu AI introduced GLM-Legal, a domain-specific legal model built on the lightweight GLM-5.3-Flash foundation. It achieved a 19.2% all-pass rate on Harvey LAB across 24 practice areas, and matched frontier closed-source models on LegalBench and PLawBench despite a leaner parameter count.
Training used mid-training on legal corpora, then post-training with reinforcement learning in realistic legal environments co-optimized with the AutoClaw harness. The model is available via API and the AutoClaw desktop agent platform. Examples show improved contract review, evidence analysis, and compliance research over the base model.
What commenters are saying
Commenters split over AI accountability in legal practice. Many worried about hallucinated cases escaping human review and lawyers not disclosing AI use to clients, with one sharing experience of paying $500/month for obvious AI-generated updates. Several argued liability still rests on the signing attorney.
Others defended AI tools as faster than manual Westlaw research, or asked whether proprietary Chinese models would gain traction. A few noted professional regulation could protect lawyers from commoditization, while others predicted a race to the bottom on document signing.