GLM-Legal Hits 19.17% on Harvey Legal Benchmark

Z.ai put a number on the table this week: 19.17% on Harvey's Legal Agent Benchmark, good for seventh place out of 79 models tracked by Vals AI. The model behind it, GLM-Legal, is built on Z.ai's GLM-5.3-Flash, a 320-billion-parameter mixture-of-experts system with 18 billion active parameters and a 1-million-token context window. Z.ai hasn't published a separate parameter count for the legal variant itself, and access runs through an application-only API with no public weights, which means outside developers can't self-host or inspect it the way they could the base model.

The headline figure matters because it's nearly three times GLM-5.3-Flash's own 6.7% baseline on the same benchmark, a jump Z.ai says comes from training aimed specifically at contract review, evidence analysis and compliance research rather than general-purpose reasoning. Harvey's LAB is built to test exactly that kind of agentic legal work, task by task, pass or fail, with no partial credit. At $0.15 per test run, the pitch is frontier-level legal performance without frontier pricing. Whether that holds up under scrutiny, and where the model actually falls short, is the harder question.