Z.ai unveiled GLM-Legal, a GLM-5.3-Flash legal agent that matches frontier models on Harvey, LegalBench, and PLawBench.
Z.ai announced GLM-Legal, a domain model built on GLM-5.3-Flash for agentic legal work such as contract review, case preparation, and document drafting. On Harvey LAB it scored a 19.2% all-pass rate across 24 practice areas, 12.5 points above its base model, and the company says it matches frontier closed models on LegalBench and Chinese PLawBench with a leaner parameter budget. Training combines expanded legal mid-training with reinforcement learning over tens of thousands of legal environments, co-evolved with an agent harness. The model is planned for AutoClaw, Z.ai's desktop agent platform.
GLM-Legal is built on the lightweight GLM-5.3-Flash base.
Harvey LAB all-pass rate is 19.2%, up 12.5 points.
It claims frontier parity on LegalBench and Chinese PLawBench.
Training uses legal mid-training plus reinforcement learning with an agent harness.
The model is slated to ship in the AutoClaw desktop agent.
Large language models are shifting from single-turn code generation toward broader professional collaboration (Professional Cowork). Real-world knowledge work introduces distinct requirements: models must digest fragmented source files under concrete objectives and tight delivery constraints, structure specialized knowledge into cohesive analysis, and maintain domain rigor across the full execution lifecycle. Legal work concentrates on these challenges, providing an ideal stress test for training models to operate in high-stakes professional environments.
Legal practice demands an unbroken chain between facts, statutory doctrine, analytical reasoning, and client-ready work product. When 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 is the true benchmark for domain-specific models.
We evaluated GLM-Legal across three axes: complex task execution, specialized legal reasoning, and real-world legal workflows. GLM-Legal achieved a 19.2% all-pass rate on Harvey LAB, while matching frontier closed-source models on LegalBench and PLawBench despite running on a significantly leaner parameter footprint.
Benchmark Highlights
Across complex execution, legal reasoning, and real-world drafting, GLM-Legal matches frontier-tier performance while maintaining high inference efficiency:
Harvey LAB (19.2% All-Pass Rate): In real-world law firm workflows across 24 practice areas, GLM-Legal scores 19.2% and improves on its base model by 12.5 percentage points. Under Harvey's strict all-pass grading, this confirms reliable end-to-end task execution without intermediate degradation.
LegalBench (Domain Reasoning Parity): 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.
PLawBench (Chinese Localized Practice): Achieves parity with leading frontier systems on Chinese legal consultation, case analysis, and document drafting under fine-grained rubrics.
Training Methodology
GLM-Legal prioritizes training efficiency. Built on the lightweight GLM-5.3-Flash foundation, it scales domain competence through specialized mid-training coupled with high-efficiency post-training. By tying foundational pre-training directly to downstream task execution, we train the model to parse source records, apply doctrine, structure arguments, and maintain output quality through final delivery.
1. Mid-training: Solidifying Domain Foundations
Mid-training establishes a systematic legal foundation. We expanded the volume of high-quality legal corpora, allowing the model to learn legal concepts, statutory rules, and interpretive frameworks within dense domain context. This stage strengthens deductive connections between facts, rules, and conclusions, providing the knowledge substrate required for complex downstream execution.
2. Post-training: Co-evolving the Model and Agent Harness
Post-training translates domain knowledge into task execution via reinforcement learning across tens of thousands of realistic legal environments. The model learns to plan analysis, advance workflows, and generate deliverables under concrete constraints.
In this setup, the Harness serves as the runtime orchestration framework. It manages task context, connects toolsets, and routes execution feedback:
Co-iterative optimization: Harness architecture and model training iterate in lockstep. Task structuring, context ingestion, and runtime feedback loops collectively govern agent exploration.
Bi-directional learning: Edge cases surfaced during model training pinpoint architectural bottlenecks in the harness. Conversely, harness refinements expose the model to realistic execution constraints, training it to retrieve relevant context, adapt via intermediate execution signals, and meet production-grade standards from prompt ingestion to final sign-off.
3. Production Feedback Loops
Real-world workflows and practitioner feedback set our optimization roadmap. In production, legal professionals evaluate outputs against concrete criteria:
Are all critical issues and edge cases addressed?
Are arguments well-substantiated and citations legally sound?
Does the deliverable strictly adhere to the matter brief?
Can downstream teams use the output immediately without manual rework?
These requirements turn subjective quality into concrete engineering loss signals. When diagnosing issues such as missed facts, shallow analysis, or format deviations, we inspect the entire execution pipeline, isolating the contributions of raw model intelligence, context staging, and workflow logic.
We feed these findings directly back into task environments, evaluation rubrics, and harness design for subsequent training cycles, validating improvements against production benchmarks. This tight feedback loop between practical requirements, task environments, the harness, and model training ensures algorithmic progress translates directly into production reliability.
Bringing GLM-Legal to AutoClaw
You’ll soon be able to put these capabilities to work in AutoClaw, our desktop-native agent platform. The examples below illustrate how GLM-Legal can help with three common legal tasks: reviewing a contract, preparing a case, and planning a cross-border launch.
1. Contract Review: Making the Exit Clause Work for the Client
Task: Review a software services contract for a client switching provider, focusing on termination and handover.
GLM-5.3-Flash
Termination and handover
The contract allows the client to terminate on 60 days’ notice and provides a 30-day cure period for material breaches. Consider shortening the notice period to give the client more flexibility. On termination, require the provider to return client data and completed deliverables for which the client has paid. Add a duty to cooperate with the replacement provider, including reasonable access to documentation and assistance with the handover. Confirm outstanding payments and any charges for transition support.
GLM-Legal
Termination and handover
Reduce the client’s convenience-termination notice from 60 to 30 days. Remove the cure period for breaches of IP, confidentiality, or data-security obligations so the client can terminate immediately. Section 4 already assigns the work product to the client: require return of all deliverables and work in progress, whether or not paid for. Add up to 90 days of transition assistance covering source code, documentation, data, and knowledge transfer to the client or its successor provider, at the provider’s then-current rates.
When changing providers, your team needs an exit clause that protects the handover.
GLM-Legal turns that goal into specific terms for immediate termination after serious breaches, return of work product, and up to 90 days of transition support. AutoClaw puts those edits into the Word document with tracked changes and comments, ready for your team to review and negotiate.
2. Evidence Analysis: Connecting the Documents to a Case Theory
Task: Analyze 40 documents from a software project dispute to identify contradictions in the provider’s progress reports.
GLM-5.3-Flash
External reports and internal records
The September 30 client report states that the project is on track and 60% complete (BR-000024), while the internal snapshot records approximately 41% completion (BR-000027). The internal emails also discuss architecture risks and instructions about client communications (BR-000014; BR-000015). These documents may support the client’s position that progress was overstated. Review the reporting history, invoicing records, and contract terms to determine whether the statements were misleading and what further evidence is needed.
GLM-Legal
External reports and internal records
The September 30 “ON TRACK – 60% COMPLETE” report (BR-000024) should be read against the September 8 warning that the milestone could not honestly be met (BR-000014) and the internal ~41% snapshot (BR-000027). The CEO’s instruction to suppress the risk and continue invoicing (BR-000015) links that internal knowledge to a disclosure decision. MSA §§2 and 5 state that payment does not waive rights and warranties survive payment (BR-000001), so continued payment alone does not resolve the dispute. These links support a concealment theory; reconcile the completion measures and test the remaining claim elements before drawing a final conclusion.
In a project dispute, the key is to connect the progress reports with what the provider knew internally.
GLM-Legal pairs external statements with conflicting records, cites the source documents, and explains their potential legal significance. The contradiction map helps your team assess the evidence and decide what to investigate next, while keeping facts and inferences distinct.
Task: Plan privacy compliance for a website launching in Singapore, Malaysia, and Indonesia, with tracking, email marketing, and overseas cloud storage.
GLM-5.3-Flash
Pre-launch priorities
Before launch, publish a privacy notice explaining collection, use, sharing, and overseas storage. Review whether cookies and analytics need consent, include an unsubscribe option in marketing emails, and ensure vendors provide suitable data-protection safeguards. Check the applicable privacy laws in Singapore, Malaysia, and Indonesia and confirm any local requirements. The launch plan should address notices, consent, marketing, and transfers, with responsibilities assigned to the relevant teams and outstanding legal questions reviewed before the website goes live.
GLM-Legal
Pre-launch priorities
P0 — recommended before launch: Publish English and Indonesian privacy notices covering purposes, transfers, rights, and complaints. Use opt-in for non-essential analytics and advertising, keep consent records, and separate marketing permission from registration. Make unsubscribe requests feed a suppression list. Map cloud, payment, email, analytics, and CDN vendors to their data destinations, then confirm the contracts and transfer safeguards required in each market (SG PDPA s26; MY PDPA s129; ID PDP Law Arts 55–56). Use these shared controls as a conservative launch baseline, while checking each market’s specific requirements.
A cross-border launch needs a plan your team can act on.
GLM-Legal breaks privacy requirements into practical steps for tracking consent, marketing permissions, and overseas data transfers, with legal references for each market. The prioritised checklist helps your team implement shared controls and identify country-specific questions to confirm before launch.
Join the GLM-Legal Beta
Whether you are a law firm, a legal tech company, an in-house legal team, or an independent practitioner, we invite you to experience GLM-Legal.
Consultation
Consultation
AutoClaw