Z.ai Opens Access to 320B GLM-Legal AI Model

Z.ai has introduced GLM-Legal, a specialized artificial intelligence model built on top of GLM-5.3-Flash, targeting law firms, in-house legal teams, and compliance officers. The underlying Mixture-of-Experts architecture features 320 billion total parameters, activating 18 billion parameters per token via 8 routed experts and one shared expert out of 288. It provides a 1-million-token context window and is currently accessible via a gated API application form or within Z.ai's AutoClaw desktop agent, which offers 200 million promotional tokens valued at $24 upon signup.

While Z.ai has not released domain-specific evaluations for GLM-Legal, independent tests from Vals AI show the GLM-5.3-Flash base model achieving an 83.9% score on LegalBench and 45.2% on Legal Research Bench. On Artificial Analysis's Intelligence Index v4.1.1, the base model scored 57 with an estimated task cost of $0.045. However, base performance on multi-step legal agent tasks like those in the Harvey benchmark indicates potential weaknesses in complex workflows that require tool usage and multi-step planning.

For legal practitioners and developers, GLM-Legal offers an alternative to domain competitors like Harvey and Thomson Reuters CoCounsel. The 1-million-token context allows teams to process entire contract sets, policy documents, and regulatory filings without chunking. However, integrating GLM-Legal into production requires careful testing around prompt reliability across long contexts, citation fidelity, and specific jurisdictional coverage, as Z.ai has yet to clarify whether the model focuses on Chinese civil law or international legal frameworks. Furthermore, teams must evaluate data privacy, API pricing, and security controls before moving beyond initial desktop testing.