Overview
GLM-Legal is a legal-domain agent model that Z.ai announced on 12 October 2026 in a post titled "GLM-Legal: Advancing Agentic Legal Intelligence". It is not a new base model: Z.ai built it on GLM-5.3-Flash, its lightweight GLM-5 series model, and specialised it for legal work. Z.ai frames legal practice as a stress test for professional agents, because the work demands an unbroken chain from source documents and statutory doctrine through analytical reasoning to client-ready work product, while meeting the exact specification of the task.
Training ran in three stages. Mid-training expanded the volume of high-quality legal corpora so the model learns legal concepts, statutory rules and interpretive frameworks in dense domain context. Post-training used reinforcement learning across tens of thousands of realistic legal environments, with the model and its agent harness — the runtime that manages task context, connects tools and routes execution feedback — iterated together. Feedback from legal practitioners in production then set the optimisation roadmap: whether critical issues are covered, whether arguments and citations are sound, whether the deliverable follows the brief, and whether it can be used without rework.
Z.ai evaluated the model on complex task execution, specialised legal reasoning and real-world legal workflows. On Harvey's Legal Agent Benchmark (Harvey LAB), which covers law-firm workflows across 24 practice areas under strict all-pass grading, Z.ai reports a 19.2% all-pass rate, 12.5 percentage points above GLM-5.3-Flash. On LegalBench (86.2) and the Chinese-practice PLawBench (77.3), Z.ai reports results on par with the frontier closed-source models in its chart. Z.ai states that the Harvey LAB and LegalBench results were independently evaluated by Vals.
The launch post shows three workflows, each set against GLM-5.3-Flash's answer to the same task: reviewing a software services contract for a client switching provider, with the edits delivered as tracked changes and comments in a Word document; analysing 40 documents from a software project dispute to map contradictions between a provider's progress reports and its internal records; and planning privacy compliance for a website launching in Singapore, Malaysia and Indonesia. Z.ai said at launch that these capabilities were coming to AutoClaw, its desktop-native agent platform.
Access at launch was a gated beta. Z.ai invited law firms, legal tech companies, in-house legal teams and independent practitioners to apply for API access through a consultation form. Z.ai published no weights, no parameter count of its own for GLM-Legal and no API pricing in the announcement.
| Released | 2026-10-12 |
|---|---|
| License | Proprietary (gated beta; no weights published) |
| Weights | API only |
| Architecture | Built on GLM-5.3-Flash, specialised through legal-domain mid-training and reinforcement-learning post-training |
| Modalities | Text |
| Status | Gated beta since 12 October 2026 — API access by application |
Benchmarks
Z.ai's "Legal Benchmark Evaluation" chart from the GLM-Legal launch post, transcribed.
| Benchmark | GLM-Legal | GLM-5.3-Flash | Kimi K3 | Claude Fable 5 | GPT-6 Astra |
|---|---|---|---|---|---|
| Harvey's Legal Agent Benchmark (all-pass) | 19.2% | 6.7% | 12.9% | 11.3% | 5.4% |
| LegalBench | 86.2 | 83.9 | 86.2 | 88.6 | 87.1 |
| PLawBench | 77.3 | 73.8 | 76.2 | 77.7 | 69.1 |
This model's scores
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- 19.2% all-pass on Harvey's Legal Agent Benchmark across 24 practice areas, 12.5 points above the GLM-5.3-Flash base
- LegalBench of 86.2, against 88.6 for Claude Fable 5 and 87.1 for GPT-6 Astra in Z.ai's chart
- PLawBench of 77.3 on Chinese legal consultation, case analysis and drafting, against 77.7 for Claude Fable 5
- Reinforcement-trained across tens of thousands of realistic legal environments rather than only fine-tuned on legal text
- Model and agent harness trained together, so the model learns to retrieve context and act on intermediate execution feedback
- Built on the lightweight GLM-5.3-Flash, with Harvey LAB and LegalBench results independently evaluated by Vals
Best for
- Contract review: turning a client's goal into specific redlines, delivered as tracked changes and comments in a Word document
- Evidence analysis: mapping contradictions across a document set, with each point cited to its source document
- Cross-border compliance research: a prioritised pre-launch checklist with statutory references for each market
- Legal tech products and in-house legal teams that need an agent model specialised for end-to-end legal deliverables
FAQ
What is GLM-Legal built on?
Z.ai built GLM-Legal on GLM-5.3-Flash. It added legal-domain mid-training on an expanded volume of high-quality legal corpora, then reinforcement-learning post-training across tens of thousands of realistic legal environments, with the agent harness and the model iterated together.
How does GLM-Legal score on Harvey's Legal Agent Benchmark?
Z.ai reports a 19.2% all-pass rate across 24 practice areas, 12.5 percentage points above the 6.7 it charts for GLM-5.3-Flash. In the same chart Kimi K3 scores 12.9, Claude Fable 5 11.3 and GPT-6 Astra 5.4. Z.ai states that the Harvey LAB and LegalBench results were independently evaluated by Vals.
How can I get access to GLM-Legal?
GLM-Legal launched on 12 October 2026 as a gated beta. Law firms, legal tech companies, in-house legal teams and independent practitioners apply for API access through Z.ai's consultation form at z.ai/consultation/legal. Z.ai also said the model was coming to AutoClaw, its desktop agent platform.
Are GLM-Legal's weights open?
No weights were published with the announcement; GLM-Legal is reachable only through Z.ai's gated beta. Its base model, GLM-5.3-Flash, is a separate release with its own published weights.