Z.ai introduced GLM-Legal, a legal-domain agentic model built on its GLM-5.3-Flash foundation, in an October 12, 2026 announcement, reporting a 19.2% all-pass rate on the Harvey LAB benchmark of real-world law firm workflows.
The company frames legal work as an ideal stress test for training models to operate in high-stakes professional environments. Legal practice, Z.ai argues, demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product: given a professional task, an agent must extract core issues from source documents, build defensible arguments, and format its findings to exact task specifications. Simultaneously satisfying all of these constraints, the company writes, is the true benchmark for domain-specific models.
Benchmark Results
Z.ai says it evaluated GLM-Legal across three axes: complex task execution, specialized legal reasoning, and real-world legal workflows.
On Harvey LAB, which grades real-world law firm workflows across 24 practice areas, GLM-Legal achieved a 19.2% all-pass rate, an improvement of 12.5 percentage points over its base model, the company reports. Z.ai says that under Harvey’s strict all-pass grading, the result confirms reliable end-to-end task execution without intermediate degradation.
On LegalBench, the company says the model parses specialized contracts, merger filings, and judicial opinions with the precision of top closed-source frontier models despite operating on a significantly leaner parameter budget. Z.ai also reports parity with leading frontier systems on PLawBench, which evaluates Chinese legal consultation, case analysis, and document drafting under fine-grained rubrics.
Training Methodology
GLM-Legal is built on what Z.ai describes as the lightweight GLM-5.3-Flash foundation, scaling domain competence through specialized mid-training coupled with high-efficiency post-training.
In the mid-training stage, Z.ai expanded the volume of high-quality legal corpora so the model learns legal concepts, statutory rules, and interpretive frameworks within dense domain context. According to the company, this stage strengthens deductive connections between facts, rules, and conclusions, providing the knowledge substrate required for complex downstream execution.
Post-training translates that domain knowledge into task execution through reinforcement learning across tens of thousands of realistic legal environments, where the model learns to plan analysis, advance workflows, and generate deliverables under concrete constraints. Z.ai describes the Harness as the runtime orchestration framework in this setup: it manages task context, connects toolsets, and routes execution feedback. Harness architecture and model training iterate in lockstep, the company says, and the learning runs in both directions: edge cases surfaced during model training pinpoint architectural bottlenecks in the harness, while harness refinements expose the model to realistic execution constraints from prompt ingestion to final sign-off.
A third stage converts production feedback into training signal. Z.ai says legal professionals evaluate outputs against concrete criteria: whether all critical issues and edge cases are addressed, whether arguments are well-substantiated and citations legally sound, whether the deliverable strictly adheres to the matter brief, and whether downstream teams can use the output immediately without manual rework. These requirements turn subjective quality into concrete engineering loss signals, which the company feeds back into task environments, evaluation rubrics, and harness design for subsequent training cycles, validating improvements against production benchmarks.
Three Legal Tasks in AutoClaw
Z.ai says users will soon be able to put GLM-Legal’s capabilities to work in AutoClaw, its desktop-native agent platform, and published three sample tasks to illustrate the integration. Each example is presented alongside output from the base GLM-5.3-Flash model for comparison.
In a contract review example, the model was asked to examine a software services contract for a client switching providers, focusing on termination and handover. The sample output proposed reducing the client’s convenience-termination notice from 60 to 30 days, removing the cure period for breaches of intellectual property, confidentiality, or data-security obligations so the client could terminate immediately, requiring the return of all deliverables and work in progress whether or not they had been paid for, and adding up to 90 days of transition assistance covering source code, documentation, data, and knowledge transfer. AutoClaw then places those edits into the Word document as tracked changes and comments.
In an evidence analysis example spanning 40 documents from a software project dispute, the sample output read a September 30 client report stating the project was on track and 60% complete against a September 8 internal warning that the milestone could not honestly be met and an internal snapshot recording roughly 41% completion. It connected those records to a CEO instruction to suppress the risk and continue invoicing, and cited sections 2 and 5 of the master services agreement, which provide that payment does not waive rights and that warranties survive payment, in support of a concealment theory, while noting the remaining claim elements should be tested before a final conclusion.
In a compliance research example, the model was asked to plan privacy compliance for a website launching in Singapore, Malaysia, and Indonesia with tracking, email marketing, and overseas cloud storage. The sample output produced a pre-launch priority list: publish English and Indonesian privacy notices, apply opt-in consent to non-essential analytics and advertising while keeping consent records, route unsubscribe requests to a suppression list, map cloud, payment, email, analytics, and content delivery vendors to their data destinations, and confirm the contracts and transfer safeguards required in each market, with statutory references to Singapore and Malaysia’s personal data protection acts and Indonesia’s personal data protection law.
Beta Access
Z.ai is inviting law firms, legal tech companies, in-house legal teams, and independent practitioners to join the GLM-Legal beta, with API access handled through a consultation application.
AutoClaw itself is a locally installed desktop application. According to AutoClaw’s product page, it supports Windows 10 and later, macOS on Apple Silicon and Intel machines, iOS, and Android, with Linux builds also offered. The app includes more than 50 built-in skills, supports switching between GLM and DeepSeek models, and connects to WhatsApp, Telegram, Discord, Lark, and other messaging platforms so users can assign tasks and receive results inside chat threads.