Z.ai Trains a Legal Model by Co-Evolving the Harness. 19.2 Percent All-Pass on Harvey LAB.
Z.ai released GLM-Legal on Monday, a legal specialist built on the lightweight GLM-5.3-Flash, and the post hit 366 points on Hacker News. The headline number is 19.2 percent all-pass on Harvey's Legal Agent Benchmark, the strict grading where a deliverable counts only if every criterion passes, up 12.5 points over the base model, with parity against the closed frontier on LegalBench and on PLawBench for Chinese practice. Meta's Muse Wonder card quotes the same Harvey LAB on the same day, which tells you legal work product has become a standard frontier scoreboard.
The interesting part is the method, not the score. Z.ai describes mid-training on a larger legal corpus, then RL across tens of thousands of simulated legal environments, and it says in plain words that the harness and the model were trained in lockstep: edge cases surfaced in training exposed bottlenecks in the harness, harness changes exposed the model to realistic execution constraints, and production feedback from practitioners, did it miss an issue, is the citation sound, can the next team use it without rework, was turned into loss signals. The harness is treated as a trainable component, not a wrapper. That is the harness-not-the-model thesis stated by a lab as its training recipe, in a domain where the harness has to open a Word file and leave tracked changes.
The examples make the difference legible. Asked to review an exit clause, the base model suggests shortening the notice period. GLM-Legal cuts it from 60 to 30 days, removes the cure period for IP and data-security breaches so the client can terminate immediately, and adds 90 days of transition assistance at the provider's current rates, then AutoClaw writes it into the document with comments. On a 40-document dispute, it reads the September 30 on-track report against the September 8 internal warning that the milestone could not honestly be met. The outputs look like a junior associate's work product, not a summary.
The caveat is the one that applies to every vertical model this year: the benchmark Z.ai leads on is the one Harvey built, and the deployment is Z.ai's own AutoClaw desktop agent, coming soon. Nineteen percent all-pass also means four of five matters still fail at least one criterion. But a Flash-sized model matching the frontier on legal reasoning, with the harness treated as part of the model, is the cheapest route anyone has shown to a professional-grade agent.
Blog: https://z.ai/blog/glm-legal
← Back to all articles
The interesting part is the method, not the score. Z.ai describes mid-training on a larger legal corpus, then RL across tens of thousands of simulated legal environments, and it says in plain words that the harness and the model were trained in lockstep: edge cases surfaced in training exposed bottlenecks in the harness, harness changes exposed the model to realistic execution constraints, and production feedback from practitioners, did it miss an issue, is the citation sound, can the next team use it without rework, was turned into loss signals. The harness is treated as a trainable component, not a wrapper. That is the harness-not-the-model thesis stated by a lab as its training recipe, in a domain where the harness has to open a Word file and leave tracked changes.
The examples make the difference legible. Asked to review an exit clause, the base model suggests shortening the notice period. GLM-Legal cuts it from 60 to 30 days, removes the cure period for IP and data-security breaches so the client can terminate immediately, and adds 90 days of transition assistance at the provider's current rates, then AutoClaw writes it into the document with comments. On a 40-document dispute, it reads the September 30 on-track report against the September 8 internal warning that the milestone could not honestly be met. The outputs look like a junior associate's work product, not a summary.
The caveat is the one that applies to every vertical model this year: the benchmark Z.ai leads on is the one Harvey built, and the deployment is Z.ai's own AutoClaw desktop agent, coming soon. Nineteen percent all-pass also means four of five matters still fail at least one criterion. But a Flash-sized model matching the frontier on legal reasoning, with the harness treated as part of the model, is the cheapest route anyone has shown to a professional-grade agent.
Blog: https://z.ai/blog/glm-legal
Comments