GLM-Legal matches frontier models on legal benchmarks with a leaner footprint
GLM-Legal: Advancing Agentic Legal Intelligence
GLM-Legal, built on the lightweight GLM-5.3-Flash foundation, achieves a 19.2% all-pass rate on Harvey LAB and matches top closed-source models on LegalBench and PLawBench. Its training combines legal mid-training with reinforcement learning across tens of thousands of realistic environments, co-evolving the model and agent harness. Production feedback loops turn practitioner criteria into engineering loss signals. The model will soon be available in AutoClaw for contract review, evidence analysis, and compliance research.
In real-world law firm workflows across 24 practice areas, GLM-Legal scores 19.2% and improves on its base model by 12.5 percentage points.