Key Notes
- Z.ai’s GLM-Legal is built on GLM-5.3-Flash and targets contract review, evidence analysis and compliance research that connects source documents with usable legal work.
- Vals AI reports a 19.17% all-pass score on Harvey’s Legal Agent Benchmark, where completing a task requires satisfying every criterion rather than only most requirements.
- API access is available by application and AutoClaw integration is planned, while the announcement does not specify support for Russian law or a general release date.
Z.ai has introduced GLM-Legal, a specialized AI model designed to review contracts, analyze evidence and research compliance requirements across jurisdictions. Built on GLM-5.3-Flash, the model aims to connect source documents with legal reasoning and produce work that a professional team can review and use.
The October 12 announcement opens a beta program for law firms, legal technology companies, in-house teams and independent practitioners. API access requires an application, while integration into Z.ai’s AutoClaw desktop agent is still described as forthcoming.
From Contract Clauses to Conflicting Records
In a contract-review demonstration, GLM-Legal proposes specific changes to termination and handover provisions, rather than simply identifying clauses to examine. The example connects the client’s objective of changing software providers with notice periods, serious breaches, ownership of work and transition assistance. Z.ai shows how AutoClaw could turn that analysis into tracked changes and comments in a Word document.
Another demonstration uses 40 documents from a software-project dispute. The model compares external progress reports with internal records, cites the relevant files and explains how conflicting accounts might affect a case theory. Z.ai presents the output as a map for further investigation, with factual observations separated from legal inferences.
That connection between evidence and a usable work product is the broader challenge facing legal AI. A document can contain an apparent contradiction without establishing why it occurred or what follows from it. Useful assistance requires preserving those distinctions while giving reviewers a clear route back to the supporting material.
A 19.2% Benchmark Score With Strict Grading
Z.ai reports a 19.2% all-pass rate on Harvey’s Legal Agent Benchmark, an improvement of 12.5 percentage points over the base model. Independent evaluator Vals AI lists GLM-Legal at 19.17% on that test, alongside 86.23% on LegalBench and 41.83% on Legal Research Bench.
The all-pass measure requires a task to satisfy every grading criterion before it counts as completed. Vals’s benchmark explanation distinguishes that result from the share of individual criteria met: its published breakdown puts GLM-Legal at roughly 92.5% on individual criteria, while about 19.2% of tasks pass in full.
Those figures describe different levels of completion, so the headline score should not be read as a general legal-accuracy percentage. An output can satisfy most requirements and still miss something that prevents the entire assignment from passing. Conversely, clearing a benchmark task does not establish dependable performance on every jurisdiction or type of matter.
The launch joins a wider push into professional AI, including Astra for Law and the legal-agent capabilities discussed in our Muse Wonder coverage. Their published evaluations use different setups, making the details of each test important when comparing competing claims.
Beta Access and Jurisdiction Coverage
Z.ai’s cross-border example examines privacy compliance for a website launching in Singapore, Malaysia and Indonesia. Its PLawBench discussion also addresses Chinese legal consultation, case analysis and drafting. These examples demonstrate the intended scope, but the launch does not provide a comprehensive list of supported legal systems or specify support for Russian law.
Teams can request access through the company’s application form, which asks about their organization and intended use. The beta announcement does not set out a general release date or a complete commercial pricing schedule. Vals publishes costs for its evaluation runs, but those figures do not establish the terms an organization would receive through the beta program.
For firms assessing the model, the meaningful question is how much checked, usable work it produces on their own documents. Citations and contradiction maps can make review easier, but their value depends on whether the underlying sources support the analysis and whether the relevant law is current. The beta gives Z.ai a route to test that fit with practitioners before a broader rollout.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
