Z.ai unveils GLM-Legal, an AI model for legal work that is only in beta for now
IA en un minuto newsroom · Editor: Jon Elgezabal
In 30 seconds
Z.ai has opened a beta of GLM-Legal, a version of its lightweight GLM-5.3-Flash model tuned for lawyers' work, reachable by applying to use the API. The company claims that on Harvey LAB it completes 19.2% of law firm assignments, ahead of Kimi K3, Claude Fable 5 and GPT-6 Astra. The numbers come from Z.ai itself, which has shared neither pricing nor a general release date.
Z.ai, the lab behind the GLM models, also known as Zhipu AI, has unveiled GLM-Legal, an artificial intelligence model specialized in legal work. It announced it on its blog on Monday, October 12. For now it is in beta, and access to the API, the route for connecting the model to other software, has to be requested.
It is built on GLM-5.3-Flash, the company's lightweight model. Z.ai says it first trained it on a larger body of legal texts and then on tens of thousands of environments that reproduce real legal assignments.
The company measured it with three tests. The first is Harvey LAB, made up of real law firm tasks in 24 practice areas and graded strictly: an assignment only counts if it passes every check. There GLM-Legal scores 19.2%, which is 12.5 points above its base model. In Z.ai's chart, three models from other labs trail it: Kimi K3 at 12.9%, Claude Fable 5 at 11.3% and GPT-6 Astra at 5.4%.
On the other two, the company maintains that it matches the most advanced closed models with a smaller model. On LegalBench, which tests legal reasoning, it scores 86.2, against 88.6 for Claude Fable 5 and 87.1 for GPT-6 Astra. On PLawBench, which covers Chinese legal practice, it scores 77.3, against 77.7 for Claude Fable 5.
As examples, Z.ai shows the review of a software services contract, the analysis of 40 documents from a dispute and a privacy plan for launching a website in Singapore, Malaysia and Indonesia. It also says the model will soon reach AutoClaw, its agent platform for the desktop, without giving a date.
The beta is aimed at law firms, legal tech companies, in-house legal teams and independent practitioners. The results are Z.ai's own measurements, and the announcement includes neither a price nor a date for general release.
Why it matters · analysis and opinion
Z.ai is not presenting a bigger model here, but a lightweight one trained in depth for a single trade. That is the bet worth watching: if a small, specialized model performs like the most advanced general ones on legal work, the advantage shifts to cost and to how easy it is to put to work. Real use is still some way off. The tests were run by the company itself; even its best result leaves the large majority of complete assignments short of a pass; and there is no price and no detail on how clients' documents are handled. A law firm or an in-house legal team can request access and try it on matters that are already closed, keeping a lawyer's review over everything it produces.
Official source: Z.ai · Written with the help of AI: how we make the news
