Z.ai Launches GLM-Legal Beta: Matches Frontier Models on LegalBench and PLawBench

Z.ai has introduced GLM-Legal, a specialized model designed for agentic legal intelligence, built on the lightweight GLM-5.3-Flash foundation. The model targets high-stakes professional workflows, emphasizing the ability to digest fragmented source files, structure specialized knowledge, and maintain domain rigor across the full execution lifecycle.

Benchmark Performance

GLM-Legal was evaluated across complex task execution, specialized legal reasoning, and real-world workflows, demonstrating efficiency gains without sacrificing accuracy:

• Harvey LAB: Achieved a 19.2% all-pass rate across 24 practice areas, marking a 12.5 percentage point improvement over its base model. This score reflects reliable end-to-end task execution under strict grading criteria.
• LegalBench: Matched the precision of top closed-source frontier models in parsing specialized contracts, merger filings, and judicial opinions, despite operating on a significantly smaller parameter budget.
• PLawBench: Reached parity with leading frontier systems in Chinese legal consultation, case analysis, and document drafting under fine-grained rubrics.

Training Methodology

The model employs a two-stage training process focused on efficiency and agent-harness co-evolution:

• Mid-training: Expanded high-quality legal corpora to solidify domain foundations, strengthening deductive connections between facts, statutory rules, and conclusions.
• Post-training: Utilized reinforcement learning across tens of thousands of realistic legal environments. The Harness serves as a runtime orchestration framework, managing task context and toolsets. This stage features co-iterative optimization, where harness architecture and model training iterate in lockstep, and bi-directional learning, where edge cases identify architectural bottlenecks while harness refinements expose the model to production-grade constraints.

Production Integration and Use Cases

GLM-Legal will be integrated into AutoClaw, Z.ai’s desktop-native agent platform, to handle tasks such as contract review, evidence analysis, and compliance research. The model demonstrates capabilities in generating specific legal edits (e.g., modifying termination clauses), connecting contradictory evidence in disputes, and prioritizing cross-border privacy compliance steps.

The company is currently inviting law firms, legal tech companies, and practitioners to apply for API access and beta testing via their consultation portal.