Z.ai, the lab behind the GLM model family, has opened a beta of GLM-Legal, a model built for professional legal work, and its headline number is 19.2%. That is the share of tasks it completed with nothing failing on the Harvey Legal Agent Benchmark (Harvey LAB), a test drawn from live work at law firms, spread over 24 areas of practice.

Read the other way, roughly four attempts in five did not fully pass. The company’s own blog post, published 12 October, takes the opposite view: the score “confirms reliable end-to-end task execution.” Both readings are arithmetically true. Only one of them tells you what happens if the model drafts something that goes out under a lawyer’s name.

Strict all-pass grading exists because partial credit is worthless in a filing. A brief that gets the statute right and the deadline wrong is a wrong brief. By that standard, 19.2% is a measure of how often a human reviewer could skip the redo, and the answer is still “rarely.” Legal may be the hardest place to ship a model that is right four times in five, which is why a human in the loop is not optional here.

The gain over its starting point is real, though. Z.ai says GLM-Legal beats its base model, GLM-5.3-Flash, a lighter model, on the same test, with a margin of 12.5 percentage points. Training had two stages: more legal text up front, then reinforcement learning, which rewards good results, run in tens of thousands of lifelike legal setups, with the surrounding software tuned in step with the model.

On the other two benchmarks, the company says it matches the best closed-source frontier models. LegalBench tests reasoning over judicial opinions, merger filings and contracts. PLawBench is graded on Chinese practice: drafting documents, analysing cases and answering client-style questions. Z.ai says it gets there on “a significantly leaner parameter footprint,” but it never says how large the model is, so the efficiency claim cannot be checked against anything.

The sourcing is uneven. The post says the Harvey LAB and LegalBench results “were independently evaluated by Vals,” and it points to a page on Vals’ site for the model. No such statement accompanies PLawBench, so treat that one as Z.ai’s own number. The company chose the framing of everything else in the post.

It also chose its three demonstrations, each shown beside output from GLM-5.3-Flash. One shows what “more specific” means in drafting. The prompt was a contract for a client changing software vendors, with attention on the exit. Among the edits GLM-Legal proposed: a 60-day notice period for ending the deal without cause shrinks to 30, and there is no cure period for IP, confidentiality or data-security breaches, so the client can “terminate immediately.” It also required handing back every deliverable and unfinished piece of work “whether or not paid for,” plus as much as 90 days of handover support for code, documents and data. The point is that the advice comes with numbers attached. Those are Z.ai’s chosen examples, not test results.

The other two were a hunt for contradictions across 40 documents from a software dispute, and a privacy plan for a site going live in Indonesia, Malaysia and Singapore, citing specific statute sections in each country.

What exists today is a beta. Developers must apply for API access via a form. According to the post, the capabilities are coming to AutoClaw, the company’s desktop agent product, and phrases it as “You’ll soon be able to.” No licence, price or general availability date is given.

For a firm weighing a pilot, the useful number is the 80.8% that did not fully pass, because that is the share a reviewer would have to catch. A legal tech buyer should ask Z.ai for the failure breakdown behind it: which task types fail, and whether they fail loudly or quietly.

Based on Z.ai’s blog post “GLM-Legal: Advancing Agentic Legal Intelligence” of 12 October 2026.