Z.ai is inviting applications for GLM-Legal, a legal-task model built on GLM-5.3-Flash. Its October 12 announcement shows examples of contract editing, evidence analysis and privacy research, and reports results on three legal benchmarks. The access page asks teams to apply for API access and discuss collaboration. It does not give applicants immediate, general access or a published GLM-Legal price.

The big change

  • What changed: Z.ai is offering legal teams a route to test a model trained for legal tasks, with API access handled through an application. The public evidence consists of company examples and reported benchmarks; independent legal performance testing is still needed.
  • Why it matters: A legal team evaluating such a system can ask whether it preserves the trail from matter documents to suggested edits, cited evidence or research. The examples make those tasks concrete, while their correctness still needs checking against source material and applicable law.
  • What to watch: Beta access terms, reproducible benchmark settings and results on a team's own reviewed matters will determine whether GLM-Legal can move beyond a sales discussion. The announcement leaves those purchasing and validation questions open.

The offer: apply for API access, with AutoClaw planned

Z.ai describes GLM-Legal as a specialization of GLM-5.3-Flash. It says the training combines additional legal material with reinforcement learning in realistic task environments and a tool-connecting agent harness. The announcement does not supply enough detail to reproduce that training or isolate the contribution of the model from the harness.

The company says GLM-Legal will be usable in AutoClaw, its desktop agent platform, soon. That is a planned integration in the announcement, not evidence that the illustrated workflow is available to every AutoClaw user today. For the API, the consultation form requests a company name and size, contact details, a role and usage scenarios. It says users can apply for GLM-Legal API access and discuss potential collaboration. The public pages reviewed for this article specify no beta acceptance criteria, start date, API model identifier, technical documentation, service terms specific to GLM-Legal or price. A team needs to obtain those details directly before planning an integration.

What the examples show

Z.ai presents three side-by-side examples against its base GLM-5.3-Flash model. In a software services contract, the GLM-Legal sample proposes specific termination and handover edits. Z.ai says AutoClaw can put the edits into a Word document with tracked changes and comments for a team to review. In a software project dispute, the sample connects a progress report to other documents and identifies an inference to investigate. In a website launch across Singapore, Malaysia and Indonesia, the sample organizes privacy issues and cites laws for further checking.

These are Z.ai's selected demonstrations. The public post does not show the full input files, all intermediate tool steps, reviewers' corrections or a measured error rate for the examples. Their value for a buyer is the definition of tasks to test: can the system identify a clause accurately, trace a proposition to a particular document and keep jurisdiction-specific research current? The examples themselves do not establish that it will do so on a firm's matters.

The benchmark claims and their limits

Z.ai reports a 19.2% all-pass rate on Harvey LAB, 12.5 percentage points above its base model. Harvey's benchmark documentation defines all-pass as the share of runs that satisfy every rubric criterion. The result is therefore narrower than a claim that 19.2% of legal work is safe to deliver. Z.ai describes LAB as spanning 24 practice areas, but does not publish in this announcement the exact task subset, model and harness settings, comparison runs or scored outputs behind its number. Its statement that the result confirms reliable end-to-end execution goes further than the evidence a reader can independently inspect here.

For LegalBench, Z.ai claims parity with leading closed models on legal reasoning. For PLawBench, which it describes as covering Chinese legal consultation, case analysis and drafting, it makes a similar parity claim. The announcement gives no numerical scores, compared model versions, task selections, prompts or uncertainty intervals for either. It also makes an efficiency claim without a measured GLM-Legal cost or latency figure. The benchmark names identify areas of evaluation; the published comparisons cannot be independently reconstructed from the announcement.

An evidence checklist for legal buyers

A useful beta evaluation can begin with non-client or appropriately cleared materials and tasks whose answers a qualified reviewer already knows. For each proposed use, ask Z.ai and the testing team to record:

  1. Access and economics: the beta agreement, eligibility, API endpoint and model version, rate and context limits, pricing, and whether the proposed AutoClaw workflow is actually enabled.
  2. Document handling: where uploaded matter files and prompts are processed and retained, who can access them, whether they enter training, deletion controls, subcontractors and the terms governing confidential or privileged material. The announcement and application page do not settle these questions.
  3. Task evidence: the source file and passage behind every factual claim, contract edit and legal citation; whether a reviewer can reproduce the result from the same inputs; and how the system flags missing documents or conflicting evidence.
  4. Evaluation design: an agreed set of representative matters and jurisdictions, a fixed model and harness configuration, a relevant base model, scoring by qualified reviewers and a record of omissions and unsupported citations. Ask for the task list and rubric behind each benchmark claim before treating a headline score as comparable.
  5. Human responsibility: a defined review and sign-off step before any output reaches a client, filing, negotiation or compliance decision. The examples put a human team in that role; the announcement does not demonstrate that the model can replace it.

Z.ai has offered a beta conversation and examples worth testing. Its public evidence is still too limited to infer legal accuracy, protection of client information or readiness for unsupervised work.

Sources & further reading

  • Z.ai, “GLM-Legal: Advancing Agentic Legal Intelligence”, October 12, 2026. The first-party announcement supplies the model description, selected task examples and company-reported benchmark claims. BIG CHANGE did not run GLM-Legal or independently reproduce the results.
  • Z.ai, GLM-Legal consultation form, accessed October 13, 2026. This is the public application route for API access and collaboration discussion. It lists form fields but no GLM-Legal price or acceptance terms.
  • Harvey LAB evaluation strategies, accessed October 13, 2026. The benchmark operator explains all-pass scoring. This documentation does not verify Z.ai's reported 19.2% result.
  • Stanford's LegalBench project, accessed October 13, 2026. Background on the benchmark suite named by Z.ai, not an independent GLM-Legal evaluation.