Z.ai announces GLM-Legal, a legal-work model built on GLM-5.3-Flash
19.2% all-pass on Harvey LAB, closed beta
Partly confirmed
The takeaway
Zhipu’s Z.ai trained a domain model for legal work with reinforcement learning in ‘tens of thousands of realistic legal environments’. It is offered through an application-only API beta and the AutoClaw desktop agent, with no weights or prices published.
Status
- Claim
Partly confirmed
- Our reporting
- Medium confidence
- Importance
- 2 of 5
- Last verified
- 12 October 2026
Your AI and this story
- GPT-6 Astra165 days after its cutoff
- Claude Opus 5.5104 days after its cutoff
- Gemini 3.8 Flash195 days after its cutoff
- Grok 4.7134 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 104 days before it.
Key facts
- Source: z.ai/blog/glm-legal, published 12 Oct 2026 (page timestamp 16:08 GMT)
- Base and training: ‘Built on the lightweight GLM-5.3-Flash foundation’; mid-training on an expanded legal corpus; post-training ‘via reinforcement learning across tens of thousands of realistic legal environments’, with the agent harness and the model iterated together
- Harvey LAB: ‘GLM-Legal scores 19.2% and improves on its base model by 12.5 percentage points’ under all-pass grading, ‘in real-world law firm workflows across 24 practice areas’
- LegalBench and PLawBench: parity with ‘top closed-source frontier models’ is claimed in words; the post’s text gives no scores (a chart is shown as an image)
- Access: ‘Apply for API access’ through a consultation form, and ‘You’ll soon be able to put these capabilities to work in AutoClaw, our desktop-native agent platform’
- Not stated: parameter count, context window, pricing, API model id, open weights. A Hacker News commenter read it as ‘not an open model’
- Reach: Hacker News front page, about 300 points on 12 Oct 2026
- All benchmark figures are Z.ai’s own; no independent evaluation was found on 12 Oct
What happened
Z.ai presents GLM-Legal as a test of training a model for a profession and not only for code. The post describes three stages: mid-training on legal text to build knowledge of concepts, statutes and interpretation; reinforcement learning in simulated legal tasks, where the model plans the analysis and produces the deliverable; and a feedback loop from practising lawyers who grade outputs for missed issues, unsupported arguments and format.
Three demonstrations compare it with its base model GLM-5.3-Flash: a contract review, a case-preparation “contradiction map” that pairs a supplier’s statements with conflicting internal records, and a privacy-compliance plan for Singapore, Malaysia and Indonesia with statute references.
Why it matters
A Chinese lab known for open-weight GLM models reports results on a benchmark from the US legal-AI company Harvey and offers the model only through a closed beta. The headline number is also a reminder of where legal agents stand: under all-pass grading, the tuned model completes about one task in five.
Sources
2 sources from 2 sites. Numbers match the chips in the text.
2 sources: 1 primary, 1 reaction
Primary
Reactions
- Hacker News discussionnews.ycombinator.com, discussion
Changes
- Filed from the Z.ai blog post (read through a text mirror, the page needs JavaScript) and the Hacker News thread