On October 12, 2026, the Chinese AI company Z.ai (Zhipu AI) announced GLM-Legal, an AI model specialized for legal work. It is an "agentic" model that carries out tasks along the lines of real legal workflows, such as reviewing contracts, analyzing evidence and researching regulatory compliance. It is currently offered as a trial (beta), and the company is accepting applications for access from law firms, corporate legal departments and others.
What is GLM-Legal?
According to Z.ai, GLM-Legal is a model built by further training the company's lightweight model GLM-5.3-Flash with a focus on the legal domain. The base model GLM-5.3-Flash has 320 billion parameters in total (a figure indicating model size), of which only 18 billion are actually used in each pass, a design that saves computation. According to its developer documentation, it can handle long texts of up to 1 million tokens at once.
Z.ai explains that the role of AI is expanding from "writing code from a single instruction" to working alongside professionals. It says legal work is well suited as a subject for training AI for professional tasks, because it requires extracting key points from scattered materials, applying the law to build coherent analysis and finishing work in prescribed formats.
Evaluation results
The main results published by Z.ai are as follows.
- Harvey LAB: an all-criteria pass rate of 19.2%, up 12.5 percentage points from the base model GLM-5.3-Flash (6.7%). Among the models it compared against, Kimi K3 scored 12.9%, Claude Fable 5 11.3% and GPT-6 Astra 5.4%
- LegalBench: 86.2% on reading contracts, merger-related documents and court decisions. Z.ai says this is on par with leading closed models, but it is slightly below Claude Fable 5 (88.6%) and GPT-6 Astra (87.1%)
- PLawBench: 77.3% on Chinese legal consultation, case analysis and document drafting. Roughly the same level as Claude Fable 5 (77.7%)
Harvey LAB is an evaluation published by the legal AI company Harvey that sets tasks close to the real work of law firms and corporate legal departments across 24 practice areas. Scoring is strict, counting a task as passed only if all evaluation criteria are met, with no partial credit.
According to Z.ai, the Harvey LAB and LegalBench results were independently measured by the AI evaluation company Vals AI. PLawBench is Z.ai's own evaluation. On the Vals AI leaderboard, GLM-Legal's Harvey LAB score ranked 7th out of 79 models.
Training method
Z.ai says it carried out training in the following three stages.
- Mid-training: training on large amounts of high-quality legal text to learn legal concepts, statutes and approaches to interpretation
- Post-training: reinforcement learning (learning that improves through trial and error) in tens of thousands of realistic legal work environments, practicing everything from planning an analysis to producing deliverables
- Feedback from practice: legal professionals evaluate outputs against criteria such as "Are any important issues missed?", "Are citations accurate?" and "Can it be used as is?", and the results are fed into the next round of training
In post-training, the company says it improved the model together with the harness, the mechanism that gives the model tools and materials and has it carry out tasks.
Intended uses
In its announcement, Z.ai showed the following three use cases.
- Contract review: for a client changing vendors, reviewing the termination clauses of a software development contract and proposing specific revisions
- Evidence analysis: in a dispute over a project, reading 40 documents and organizing discrepancies in progress reports while citing the source documents
- Regulatory compliance research: for a website launching in Singapore, Malaysia and Indonesia, compiling a prioritized list of personal data protection measures
In the contract example, the company says its desktop agent AutoClaw writes revisions into a Word document in the form of tracked changes and comments.
How to use it and availability
GLM-Legal is offered by applying for API access through the application page on Z.ai's website. The company lists law firms, legal tech companies, corporate legal departments and individual practitioners as the target users. It is also scheduled to become available in AutoClaw soon. Z.ai has not published pricing.
According to Artificial Lawyer, a U.K. media outlet specializing in legal tech, the model weights (the data that makes up the model) have not been released at this time. Availability in Japan and support for Japanese law have not been announced.
CoAI's take
GLM-Legal is an example of an AI model developer releasing a dedicated legal model itself, rather than an add-on feature for a general-purpose model. The approach of training a lightweight model in a specific field to approach the performance of large models is attractive to companies that want to keep costs down. On the other hand, while some evaluations were measured by a third party, the comparisons with leading models come from Z.ai's announcement, and pricing and support for Japanese law are unknown. When using AI for legal work, it is essential to handle confidential information carefully and to put in place a system in which people ultimately check the content.