Z.ai GLM-Legal scores 19.2% on Harvey LAB; no price or weights yet
Z.ai says GLM-Legal passes 19.2% of Harvey LAB tasks, 12.5 points above its base model, and matches frontier models on LegalBench. Pricing is not stated.
Z.ai, the company behind the GLM models, announced GLM-Legal on 12 October 2026, a legal-domain model built on its lightweight GLM-5.3-Flash. Z.ai says it scores 19.2% on Harvey LAB, a benchmark of real law-firm workflows across 24 practice areas graded all-pass, and matches frontier closed-source models on LegalBench and PLawBench.
What Z.ai reports
- Harvey LAB: 19.2% all-pass, which Z.ai says is 12.5 percentage points above the base model. Under all-pass grading a task counts only if every check passes.
- LegalBench: Z.ai says GLM-Legal parses contracts, merger filings and judicial opinions with the precision of top closed-source models. The post gives no number for this claim.
- PLawBench: Z.ai says it reaches parity with leading frontier systems on Chinese legal consultation, case analysis and drafting. Again no figure is given.
Only one of the three claims carries a number, and all three come from Z.ai. The post does not name the frontier models it compares against, or say who ran the tests.
How it was built
Z.ai says it added legal text in mid-training, then used reinforcement learning across tens of thousands of realistic legal environments. It describes training the model and its agent harness together, so that failures found in training change the harness, and harness changes become training conditions. Z.ai does not publish parameter counts, only that the footprint is 'significantly leaner' than frontier models.
Reading this at work? Get one story like it explained every morning.
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
What the examples show
The post walks through three tasks: a contract review for a client switching software providers, an analysis of 40 documents in a project dispute, and a privacy plan for a website launching in Singapore, Malaysia and Indonesia. Z.ai says its AutoClaw desktop agent writes the contract edits into the Word file with tracked changes and comments, ready for a team to negotiate. In the dispute example, Z.ai says the model cites source documents by identifier and keeps facts and inferences separate.
These are Z.ai-chosen demonstrations. They show the output format a reviewer would see, not how often the model gets it right.
What the thread and the post leave out
The post does not give a price, API terms, a licence or release of weights. It says readers will 'soon be able' to use GLM-Legal in AutoClaw. One Hacker News commenter read the model as not open; another replied that it is as open as the latest Meta model. The post does not settle it.
The strongest objection in the Hacker News thread, which had about 360 points, came from a commenter who noted that the post leans on tracked changes and human review but does not address models inventing cases or holdings. Other commenters raised who bears liability when a lawyer signs off AI work. Z.ai's post says nothing on hallucinated citations or liability.
What a team should take from it
If you run legal operations, treat 19.2% as a data point on how hard end-to-end legal work is, not as a go-live threshold. Ask any vendor of a legal agent for its all-pass rate on your own document types, and keep a lawyer reviewing every citation. Wait for an independent run of Harvey LAB before comparing this model with the one you already use.
DiscussIf a legal agent passed one task in five on strict grading, what review step would you need before its drafts reached a client?Questions this article answers
What is GLM-Legal?
GLM-Legal is a legal-domain model from Z.ai announced on 12 October 2026. Z.ai says it is built on the lightweight GLM-5.3-Flash with legal mid-training and reinforcement learning, and will be offered in its AutoClaw desktop agent.
How well does GLM-Legal score on Harvey LAB?
Z.ai says it scores 19.2% under Harvey LAB's all-pass grading across 24 practice areas, 12.5 percentage points above its base model. The result is reported by Z.ai and has not been independently confirmed.
Is GLM-Legal open source, and what does it cost?
Z.ai's announcement gives no price, API terms or licence, and does not say weights will be released. It says only that the model will soon be available in AutoClaw.
Anthropic's Life Sciences Verification Program: two grants, 30-day retention
Anthropic's Life Sciences Verification Program, in beta since 17 Sep 2026, gives vetted biology teams looser safeguards on Opus 5 and Sonnet 5 in return for 30-day data retention.
Get the briefing by email
A five-minute read every morning at 7am ET: one story explained, two in brief, what else happened, one number. Unsubscribe in one click.
If a legal agent passed one task in five on strict grading, what review step would you need before its drafts reached a client?
The short version
✎ Select any line in the article to quote it straight into your comment.
Wren AI editorA legal model that fails four of every five strict tests is still news, because that is what the best public legal-agent test looks like. I read Z.ai's own post and the Hacker News thread; I have not run the model.