GLM-Legal
Z.ai's legal AI model for contract review and compliance.
About
GLM-Legal is Z.ai's legal-domain model, launched October 12, 2026 in limited beta with API access by application. It is built on the company's GLM-5.3-Flash base (320B total, 18B active parameters), then given legal-domain mid-training and reinforcement learning across tens of thousands of realistic legal environments. The headline figure is a 19.17% all-pass rate on Harvey LAB — Harvey's legal agent benchmark across 24 practice areas — independently verified by Vals AI at #7 of 79 models, within the leaderboard's margin of error of Gemini 4 Argon while costing roughly $0.15 per task against $10.55 for Argon. Vals also scored it 86.2% on LegalBench (#14 of 150). It targets contract review, evidence analysis, and compliance research, with integration into Z.ai's AutoClaw desktop agent planned. The model is not open weights.
What it can do
- Legal-domain model on the GLM-5.3-Flash base with legal RL (provider-claimed)
- 19.17% all-pass on Harvey LAB, #7 of 79 (independently verified by Vals AI)
- 86.2% on LegalBench, #14 of 150 (independently verified by Vals AI)
- Contract review, evidence analysis, compliance research (provider-claimed)
- Planned integration into Z.ai's AutoClaw desktop agent (provider-claimed)
- Roughly $0.15 per benchmark task vs $10.55 for Gemini 4 Argon (provider-reported)
Setup
- Apply for the GLM-Legal beta at z.ai.
- On approval, receive API access.
- Feed matter documents and partner-style instructions through the API.
- Review outputs with a human sign-off — ~4 in 5 complex tasks still return at least one defect.
Prerequisites
Beta application approval; API access via Z.ai.
Use cases
Works with
Trust score — 5/10
Scored 8/10 (METRICS tier, metrics-v1). Evidence: headline Harvey LAB score (19.17%) and LegalBench score (86.2%) independently verified by Vals AI on its public leaderboards — rare independent confirmation for a launch-week claim; launch covered by Artificial Lawyer and 1EyeAI; official URL verified live (200, Oct 2026). Strengths: independent benchmark verification, candid admission that weights stay closed. Docked for: (1) 19.17% all-pass means ~4 of 5 complex legal tasks return at least one defect — human sign-off is mandatory; (2) limited beta, API access by application only; (3) the PLawBench parity claim is company-stated and unchecked; (4) closed weights. License: proprietary (beta, application-only API).
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 8.0/10
- Maintenance · recency, releases, issue load
- 7.0/10
- Openness · license + docs
- 3.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Launched Oct 12, 2026 in limited beta; API access by application; not open weights
- · Built on GLM-5.3-Flash (320B total, 18B active) with legal-domain mid-training + RL
- · Harvey LAB 19.17% all-pass, #7 of 79 — independently verified by Vals AI
- · LegalBench 86.2%, #14 of 150 — independently verified by Vals AI
- · ~$0.15 per benchmark task vs $10.55 for Gemini 4 Argon (provider-reported)
- · Planned AutoClaw desktop-agent integration
- · manual review: 19.17% all-pass means ~4 of 5 complex legal tasks return at least one defect — human sign-off mandatory; PLawBench parity claim is company-stated and unchecked; closed weights
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.