Z.ai introduces GLM Legal for specialist legal workflows

Z.ai’s October 12, 2026 announcement introduces GLM Legal, a specialist model for legal work. The company invites practitioners to join a beta and says integration with its AutoClaw desktop agent is coming soon.

Built on GLM 5.3 Flash, it combines additional legal training with reinforcement learning that develops the model and its operating framework together. Z.ai reports a 19.2% all pass rate on Harvey LAB, 12.5 percentage points above its base model. Most test tasks still fall below that complete pass threshold.

The company also claims comparable performance to leading proprietary models on LegalBench and PLawBench.

What the headline score measures

Harvey introduced LAB on May 6, 2026. Its first release contained more than 1,200 tasks across 24 legal practice areas and over 75,000 expert written grading criteria. Agents receive an assignment and a set of matter documents, then produce files for review.

A task passes only when every required criterion passes. Checks can cover factual accuracy, citations, conclusions and formatting. Read that way, an all pass percentage measures completed tasks under a demanding rubric. It does not describe the proportion of individual answers that are correct.

Comparison also depends on implementation. In an August 20 research report, Harvey said its own LAB runs differed from external results because of the agent framework and judging setup. That is a practical reason to inspect evaluation settings before treating scores from different reports as interchangeable.

Because LAB works from supplied matter files, it does not by itself establish reliability for every legal research setting. Harvey says its initial 24 practice areas leave substantial parts of legal work uncovered.

The other tests ask different questions

LegalBench dates to August 20, 2023. Its original paper described 162 tasks spanning six forms of legal reasoning, assembled with legal professionals. The researchers evaluated 20 open and commercial models. It provides a vocabulary for examining particular reasoning skills, rather than a complete measure of professional work.

PLawBench, first submitted on January 23, 2026, focuses on consultation, case analysis and legal document generation. Its paper describes 850 questions across 13 practical scenarios, with roughly 12,500 expert designed rubric items. Grading uses a language model evaluator aligned with human expert judgments.

That structure makes intermediate reasoning and the handling of incomplete information important alongside the final document. The authors reported substantial weaknesses among the ten models they evaluated then. Those historical findings describe that earlier test group and cannot be reused as a current result for a newly announced model.

Our coverage of Foundry evaluation limits also examines why a benchmark score needs its measurement context.