GLM-Legal Model Matches Frontier Systems on Legal Benchmarks With Leaner Footprint

GLM-Legal, a new legal-specialized AI model, is moving beyond single-turn code generation toward broader professional collaboration. Built on the lightweight GLM-5.3-Flash foundation, the model targets the demands of real-world legal practice, where an agent must extract core issues from source documents, build defensible arguments, and format findings to exact task specifications while maintaining an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product.

Benchmark Results

Evaluated across three axes, complex task execution, specialized legal reasoning, and real-world legal workflows, GLM-Legal achieved a 19.2% all-pass rate on Harvey LAB. The benchmark spans 24 practice areas in real-world law firm workflows, and the result improves on the base model by 12.5 percentage points under Harvey's strict all-pass grading, confirming reliable end-to-end task execution without intermediate degradation.

On LegalBench, which covers specialized contracts, merger filings, and judicial opinions, the model matches top closed-source frontier models despite operating on a significantly leaner parameter budget. On PLawBench, it achieves parity with leading frontier systems on Chinese legal consultation, case analysis, and document drafting under fine-grained rubrics. Results on Harvey LAB and LegalBench were independently evaluated by Vals.

Training Methodology

Training begins with mid-training on expanded high-quality legal corpora, helping the model learn legal concepts, statutory rules, and interpretive frameworks within dense domain context and strengthening deductive connections between facts, rules, and conclusions. Post-training then applies reinforcement learning across tens of thousands of realistic legal environments, with the agent harness serving as the runtime orchestration framework that manages task context, connects toolsets, and routes execution feedback.

Harness architecture and model training iterate in lockstep. Edge cases surfaced during training pinpoint harness bottlenecks, while harness refinements expose the model to realistic execution constraints. In production, legal practitioners evaluate outputs against concrete criteria, such as whether critical issues are addressed, arguments are well-substantiated, and deliverables adhere strictly to the matter brief. These findings are fed back into task environments, evaluation rubrics, and harness design, turning subjective quality into concrete engineering loss signals.

From Benchmarks to Legal Work

The model will soon be available in AutoClaw, a desktop-native agent platform. Demonstrations show it reviewing a software services contract for termination and handover terms, analyzing 40 documents in a software project dispute to identify contradictions in a provider's progress reports, and planning privacy compliance for a website launching in Singapore, Malaysia, and Indonesia.

The GLM-Legal beta is open to law firms, legal tech companies, in-house legal teams, and independent practitioners through API access and AutoClaw.

Sources: Z.Ai