Z.ai opens gated access to GLM-Legal

Z.ai has introduced GLM-Legal, a domain-specialized version of GLM-5.3-Flash for law firms, in-house counsel, and compliance teams. The company says it can match frontier general-purpose models on mainstream legal benchmarks while focusing on contract review, evidence analysis, and compliance work. Specialized benchmark results have not been released.

Prospective users must currently request access for the API or a potential collaboration. The form asks applicants to describe intended uses across legal practice, business operations, and client services. That intake process resembles a design-partner program rather than a generally available API launch.

A large mixture-of-experts base

GLM-Legal builds on GLM-5.3-Flash, a mixture-of-experts model that activates a subset of its parameters for each token. Z.ai describes the base model as follows:

Activating 18 billion parameters per token reduces inference work relative to a dense 320-billion-parameter model. The advertised 1-million-token context window can accommodate a large contract set, amendments, policies, and supporting material in one request. Actual capacity depends on tokenization, file handling, OCR quality, and API limits. Long context can reduce retrieval and chunking work, although developers still need to test whether the model consistently finds details across the full prompt.

Benchmarks establish a baseline

Z.ai has not published evaluation results for the specialized model. Independent tests from Vals AI provide the following baseline for GLM-5.3-Flash:

The evaluations use different methods and scales, so their percentages are not directly comparable. LegalBench indicates competitive performance on bounded legal questions. The Harvey result identifies multi-step agent workflows as a weaker area in the base model. GLM-Legal may improve on that baseline, but its gains remain unverified.

Where specialization may pay off

Z.ai positions GLM-Legal for work built around recurring document patterns and structured outputs. Likely applications include:

• Reviewing contracts against an internal clause playbook
• Finding missing provisions, unusual terms, and conflicting language
• Extracting facts, dates, parties, and obligations from case files
• Comparing regulations with internal policies and controls
• Organizing evidence and constructing case timelines

Multi-step tasks that call external tools, revise plans, and preserve state across many actions deserve focused testing because of the base model’s agent benchmark result. These workflows also depend on orchestration software, tool permissions, and recovery behavior, none of which a model benchmark captures fully.

Jurisdictional coverage remains undisclosed. The GLM lineage has previously emphasized Chinese civil-law tasks, and Z.ai has not detailed the legal jurisdictions represented in GLM-Legal’s training or evaluations. US contract work, EU regulatory analysis, and other jurisdiction-specific uses therefore require separate test sets.

The intake form leaves key gaps

Z.ai has not published the technical and commercial details needed for a production integration. Open questions include:

• API design: endpoint schemas, SDKs, structured-output support, streaming, and tool-calling formats
• Service limits: rate limits, concurrency, file-size caps, latency targets, and uptime commitments
• Pricing: input, output, long-context, storage, and tool-execution charges
• Data controls: retention periods, training-data policies, regional processing, encryption, and deletion
• Security: audit certifications, tenant isolation, access logs, and administrative controls
• Deployment: hosted regions, private-cloud options, and on-premises availability
• Licensing: whether GLM-Legal carries the base model’s MIT terms

AutoClaw provides a desktop route

Z.ai is also distributing GLM-Legal through AutoClaw, its desktop AI agent. The application can operate local files, office software, browsers, messaging tools, and built-in skills from natural-language instructions. That interface offers a faster path for document-level trials than building an API integration from scratch.

New registrations receive a promotional allocation of 200 million tokens, which Z.ai values at $24. The allocation provides substantial room for testing, although production teams still need model-specific pricing, privacy terms, and service guarantees before estimating deployment costs.

Evaluation should mirror legal risk

A credible pilot should use representative documents and score outputs against lawyer-reviewed answers. The test plan should cover:

1. Clause accuracy: precision and recall for required, missing, and nonstandard provisions
2. Citation fidelity: whether every conclusion points to the correct source language
3. Long-context reliability: performance when relevant text appears at the beginning, middle, and end of large inputs
4. Jurisdiction fit: results across the governing laws and regulatory regimes used by the organization
5. Agent safety: tool permissions, approval gates, recovery from failed actions, and resistance to prompt injection in documents
6. Confidentiality: handling of privileged material, retention settings, access controls, and audit records
7. Operational cost: latency and token use on realistic document sets rather than short benchmark prompts

Low base-model costs still need product proof

Artificial Analysis gives GLM-5.3-Flash a score of 57 on Intelligence Index v4.1.1 and estimates a cost of $0.045 per task. Z.ai says comparable intelligence previously cost about ten times more. Those figures apply to the base model and do not establish GLM-Legal’s eventual API price or its cost on million-token legal workloads.

GLM-Legal will compete with Harvey, Thomson Reuters CoCounsel, and legal software built around GPT or Claude with retrieval, playbooks, and workflow controls. Its prospects depend on measurable improvements from legal post-training, reliable long-context performance, and production controls that satisfy confidentiality and compliance requirements. Until Z.ai publishes specialized evaluations and API terms, the release remains a gated access program with a capable base model and several material unknowns.