Zhipu AI has released GLM-Legal, a legal-domain model built on its GLM-5.3-Flash base, targeting contract review, evidence analysis and compliance research for professional legal work.
According to the company, the model went through mid-training on legal corpora and reinforcement learning inside tens of thousands of real legal task environments, learning material retrieval, legal analysis and document delivery. Zhipu also optimized the surrounding agent framework responsible for context management, tool calling and execution feedback.
Zhipu said GLM-Legal achieved a 19.2 percent all-item pass rate on the Harvey legal agent evaluation, 12.5 percentage points higher than the base model. Independent evaluator Vals scored it 19.17 percent in its October 12 update, placing it seventh overall and the cheapest among the top ten, with a single test costing just 0.15 dollars.
The release is being read as a sign that efficient Chinese models are moving into high-value professional services, where correctness and document quality matter more than raw conversational flair. Joint optimization of professional knowledge, task environments and execution frameworks, rather than scale alone, is what makes the difference, industry observers said.
Legal tech is one of several verticals Chinese model vendors are racing to cover, following similar domain-specific releases in finance, healthcare and software engineering, as vendors seek recurring enterprise revenue beyond commodity API sales.
Source: CITIC Construction Investment research note via Phoenix Tech, October 2026.
