On October 12, 2026, Z.ai launched GLM-Legal, a clear sign that AI is moving beyond coding assistance to become a colleague in specialized professions (Professional Cowork). Z.ai chose legal work as its proving ground because the field requires connecting facts, statutes, and reasoning seamlessly to produce work that can actually be delivered to clients.
Benchmark results worth watching
- Harvey LAB: Achieved an All-Pass rate of 19.2% on real-world tasks drawn from law firms across 24 practice areas, a 12.5-percentage-point improvement over the base model. The figure may look modest, but Harvey’s criteria are extremely strict: a task counts only if it passes every single criterion.
- LegalBench: Analyzed contracts, M&A documents, and court judgments with accuracy on par with leading closed models.
- PLawBench: Matched leading systems on Chinese legal tasks, including legal consultation, case analysis, and document drafting.
The Harvey LAB and LegalBench results were evaluated by Vals, an independent evaluator.
A cost-efficient training approach
GLM-Legal is built on GLM-5.3-Flash, a small model, and was trained in three stages:
- Mid-training: Added a high-quality legal corpus to build a foundation in legal principles and interpretation.
- Post-training: Applied reinforcement learning across tens of thousands of simulated legal environments, developing the model together with Harness, the system that controls how the agent operates.
- Production feedback: Converted feedback from practicing lawyers into signals for improving the model, such as whether key issues were fully covered, whether citations were accurate, and whether the output could be used right away.
Real-world use cases with AutoClaw
Z.ai plans to bring GLM-Legal to AutoClaw, the company’s desktop agent platform, and highlighted three example tasks:
- Contract review: Proposes specific clause revisions, such as shortening the termination notice period from 60 to 30 days and capping transition assistance at 90 days. AutoClaw then applies the edits to the Word file with tracked changes.
- Evidence analysis: Spots inconsistencies between progress reports sent to the client and internal records, cites the source documents, and clearly separates facts from inferences.
- Data privacy compliance planning: Prioritizes what must be done before launching a website in Singapore, Malaysia, and Indonesia, and identifies the relevant legal provisions in each country.
GLM-Legal is now open for beta API applications from law firms, legal tech companies, corporate legal departments, and independent legal practitioners.
What makes GLM-Legal notable is not just its scores. It shows that a small model trained for a specific domain can perform close to much larger models. In the examples Z.ai released, GLM-Legal’s answers are noticeably more specific than the base model’s and ready to use immediately. Even so, the 19.2% All-Pass rate shows that complex legal work still needs review by lawyers. For now, AI is best suited as an assistant that reduces repetitive work rather than a replacement for professional judgment.
Source: GLM-Legal