Z.ai released GLM-Legal, a domain-specialized legal model built on GLM-5.3-Flash, designed for contract review, evidence analysis, and compliance workflows. In benchmark evaluations, GLM-Legal achieved a 19.2% all-pass rate on Harvey’s Legal Agent Benchmark (LAB), compared to its base model’s 6.7% score.
The underlying GLM-5.3-Flash architecture utilizes 320 billion total parameters with 18 billion active parameters, combining sparse and linear attention to reduce long-context inference costs. Independent testing by Vals AI confirmed that GLM-Legal outperformed competing models including Kimi K3, Claude Fable 5, and GPT-6 Astra on the benchmark.
Access to GLM-Legal is currently limited to an application-only API, with plans to bring the model to Z.ai’s desktop agent application, AutoClaw. The base GLM-5.3-Flash model remains publicly accessible for developer prototyping.
Why it matters
Domain-specific post-training and harness co-evolution can yield dramatic performance gains on complex vertical agent benchmarks.
Specialized open-architecture models are directly challenging top general-purpose models on professional legal reasoning tasks.
Source: marktechpost.com



