Z.ai has unveiled GLM-Legal, a specialized version of its GLM-5.3-Flash model designed for legal firms, compliance teams, and in-house counsel. The model is built on the multimodal GLM-5 architecture, featuring 320 billion total parameters with 18 billion active parameters. It utilizes a mix of sparse and linear attention to reduce costs associated with long-context processing.
The model was evaluated on the Harvey LAB benchmark, which consists of 1,250 tasks across 24 practice areas. GLM-Legal achieved a score of 19.17%, ranking 7th out of 79 models. It outperformed Kimi K3 at 12.92%, Claude Fable 5 at 11.25%, and GPT-6 Astra at 5.42%. The base model, GLM-5.3-Flash, scored 6.67% on the same benchmark.
According to data from Vals AI as of October 12, 2026, GLM-Legal passed 92.5% of individual criteria in the Harvey LAB test. On the LegalBench benchmark, the model scored 86.23%, trailing Claude Fable 5, which reached 88.56%. Vals AI conducted these tests with two LLM judges and no internet access, using tools for docx, pptx, and xlsx files.
Z.ai has not yet disclosed pricing or a release date for GLM-Legal API access, and no repository is currently available on Hugging Face. The base model, GLM-5.3-Flash, is available for prototyping at $0.15 per 1 million input tokens and $0.50 per 1 million output tokens. The company plans to integrate GLM-Legal into its desktop agent application, AutoClaw.
