Z.ai Launches GLM-Legal Model for Legal Workflows

Z.ai announced the launch of GLM-Legal, a legal-domain agentic model, on 12 October 2026. This model is built on the GLM-5.3-Flash foundation and achieved a 19.2% all-pass rate on the Harvey LAB benchmark, which assesses real-world law firm workflows across 24 practice areas. The company claims this rate reflects a 12.5 percentage point improvement over its previous model.

Z.ai highlights that legal workflows present an ideal environment for training models due to their high-stakes nature. Legal tasks require a seamless connection between facts, statutory doctrine, analytical reasoning, and client-ready outputs. According to the company, an effective model must extract core issues from documents, formulate defensible arguments, and present findings according to strict task specifications.

The evaluation of GLM-Legal focused on three key aspects: complex task execution, specialized legal reasoning, and real-world legal workflows. On the Harvey LAB benchmark, Z.ai reports that GLM-Legal’s results confirm reliable end-to-end task execution without degradation at intermediate stages.

In addition to its performance on Harvey LAB, GLM-Legal demonstrated proficiency in parsing specialized contracts, merger filings, and judicial opinions on LegalBench. The model is said to operate with precision comparable to leading closed-source models, despite having a significantly leaner parameter budget. Z.ai also claims parity with other frontier systems on PLawBench, which assesses Chinese legal consultation and case analysis.

GLM-Legal’s architecture is based on what Z.ai describes as the lightweight GLM-5.3-Flash foundation. This model benefits from specialized mid-training that expands the volume of high-quality legal corpora. During this phase, the model learns legal concepts and statutory rules within a dense domain context, which enhances its ability to connect facts, rules, and conclusions.

The post-training phase employs reinforcement learning across numerous realistic legal environments. Here, the model learns to plan analyses, advance workflows, and generate deliverables under specific constraints. Z.ai refers to the runtime orchestration framework in this setup as the Harness. This system manages task context, connects toolsets, and routes execution feedback, ensuring that improvements in harness architecture and model training occur in tandem.

The model’s development includes a feedback loop where production feedback is converted into training signals. Legal professionals evaluate the model’s outputs against specific criteria, such as the thoroughness of issue coverage, the substantiation of arguments, and adherence to deliverable requirements. This process transforms subjective quality assessments into concrete engineering signals, which inform subsequent training cycles.

Z.ai plans to integrate GLM-Legal’s capabilities into AutoClaw, its desktop-native agent platform. The company has published sample tasks demonstrating this integration, comparing outputs from GLM-Legal with those from the base GLM-5.3-Flash model.

For instance, in a contract review scenario, GLM-Legal analyzed a software services contract for a client transitioning providers. The model recommended reducing the notice period for convenience termination from 60 days to 30 days and suggested removing the cure period for breaches related to intellectual property and data security. It also proposed requiring the return of all deliverables and work in progress, regardless of payment status, and included up to 90 days of transition assistance.

In another example involving evidence analysis from a software project dispute, the model compared various documents and identified discrepancies between client and internal reports regarding project completion. This analysis connected the findings to a CEO’s instruction for ongoing invoicing, citing relevant sections of the master services agreement to support a concealment theory.

Additionally, in a compliance research example, GLM-Legal outlined a pre-launch priority list for a website launching in Singapore, Malaysia, and Indonesia. The model suggested necessary steps for privacy compliance, such as publishing privacy notices, applying consent mechanisms, and mapping vendor data flows.

Z.ai’s GLM-Legal exemplifies advancements in legal technology, focusing on enhancing efficiency and accuracy within legal workflows. The model’s capabilities reflect ongoing developments in the integration of AI into professional environments.