The Pulse
Z.ai Launches GLM-Legal With a 19.2% Harvey LAB Pass Rate
Z.ai introduced GLM-Legal, a legal-work model built on GLM-5.3-Flash, and reported a 19.2% all-pass rate on the Harvey LAB benchmark. The company says the model will be available in beta and is coming to its AutoClaw desktop agent.
AI.info Team · Written with AI ·
GLM-Legal passed every evaluation criterion in 19.2% of its Harvey LAB runs, Z.ai says, as it introduced a model built for legal workflows rather than general-purpose chat. The all-pass score improved 12.5 percentage points over the model’s GLM-5.3-Flash base, according to the company. The result still means that most benchmark runs did not clear every criterion.
Z.ai frames the model as an agent expected to carry a task from source documents through legal analysis to a finished deliverable. Its launch post says GLM-Legal matches frontier closed-source models on LegalBench and China-focused PLawBench, while using a smaller parameter footprint. The announcement does not give exact scores for those two benchmarks or specify the model’s parameter count.
Harvey LAB tests the whole workflow
Harvey LAB evaluates law-firm workflows across 24 practice areas, according to Z.ai. The company highlights its all-pass rate as a measure of whether a system can complete a task without missing a required part along the way. That makes the 19.2% figure distinct from a score on a narrower question-answering test: a run must meet the benchmark’s criteria across the task.
The company describes GLM-Legal’s target work as extracting key issues from records, applying legal rules, building arguments and formatting results to a matter’s requirements. “Legal practice demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product,” Z.ai writes in its announcement. The post presents that chain—not isolated legal recall—as the test the model is meant to address.
Z.ai’s claims about LegalBench and PLawBench are less specific in the launch material. It says the model performs at the level of leading closed-source systems on specialized contracts, merger filings and judicial opinions, and reaches parity on Chinese legal consultation, case analysis and drafting. Without scores, named comparison models or a fuller account of the testing setup in the post, readers cannot assess those comparisons as closely as the Harvey LAB result.
GLM-5.3-Flash gets legal training and an agent harness
GLM-Legal is built on GLM-5.3-Flash. Z.ai says it expanded the model’s legal training data during mid-training, then used reinforcement learning across “tens of thousands of realistic legal environments” in post-training. The company says those stages teach the model to parse records, apply doctrine, organize arguments and produce work in a required format.
The training also pairs the model with a software harness that manages task context, connects tools and routes feedback. Z.ai says it develops that harness alongside the model: problems exposed in training can point to weaknesses in the runtime system, while changes to the harness give the model more realistic conditions for learning. The post does not identify the legal datasets or explain how the training environments were constructed.
For production feedback, the company says it tracks whether outputs address key issues, support arguments with legally sound citations, follow the matter brief and can be used without manual rework. Z.ai describes these checks as engineering signals for later training cycles. The launch post does not provide independent validation of the resulting work or detail how practitioners review it.
AutoClaw examples show contract, evidence and compliance tasks
Z.ai says GLM-Legal will soon be available through AutoClaw, its desktop agent platform, and invites law firms, legal technology companies, in-house teams and independent practitioners to join a beta. The post does not state a general release date or price.
Its sample tasks include reviewing a software-services contract, analyzing 40 documents from a project dispute, and planning privacy compliance for a website launch in Singapore, Malaysia and Indonesia. In the contract example, the model proposes changes to termination rights and a handover clause. In the dispute example, it connects a progress report, an internal completion estimate and emails, while warning that remaining claim elements need to be tested.
Those examples illustrate the workflow Z.ai is selling, but they are demonstrations selected by the company, not independent evaluations. The central open questions are whether the benchmark results hold up under outside testing, how often the agent’s legal citations withstand review, and when beta users can access it through AutoClaw.