Z.ai has unveiled GLM-Legal, a legal-domain agent model built on its GLM-5.3-Flash base, with a self-reported 19.2% all-pass rate on Harvey's Legal Agent Benchmark (Harvey LAB), 12.5 percentage points above the base model, according to Z.ai's announcement as summarized by Unite.AI. The story spread quickly on Hacker News, where the announcement passed 100 points within hours. But the headline "No. 1 open-weight legal AI" that appeared on some aggregators does not match what we could verify about access.
This post separates what Z.ai claims, what independent coverage confirms, and what you should check before trusting it with legal work.
TL;DR: claimed versus verified
| Question | Answer |
|---|---|
| What is it? | A legal-domain agent model built on GLM-5.3-Flash, in beta |
| Headline score | 19.2% all-pass on Harvey LAB, +12.5 points over base (Z.ai, self-reported) |
| Other benchmarks | No other GLM-Legal scores confirmed |
| Open weights? | Not confirmed. Access appears to be by application |
| Price | No GLM-Legal API price found |
| Base model | GLM-5.3-Flash (specs and license not confirmed by us) |
| How to try | Apply via Z.ai's consultation form, or AutoClaw, Z.ai's desktop agent |
What Z.ai says it built
According to Unite.AI's summary of the announcement, GLM-Legal is trained in stages:
- Mid-training on expanded, high-quality legal corpora to build domain knowledge.
- Post-training with reinforcement learning across tens of thousands of simulated legal environments.
- A runtime layer called the Harness that manages task context, toolsets and execution feedback.
- Feedback from legal professionals on outputs, fed back into training.
Target users are law firms, legal tech companies, in-house teams and independent practitioners. Example tasks include contract review, evidence analysis across a set of 40 documents, and privacy compliance planning for Southeast Asia.
What the Harvey LAB number means
Archive folder lifted from a tray, showing legal evidence analysis across many documents with Harvey LAB agent tasks
Harvey LAB is an agent benchmark of realistic law firm tasks in which a model must satisfy a set of rubric criteria. The "all-pass" metric means a task counts only if every criterion is met. That is deliberately harsh, which is why 19.2% sounds low but can still be a large step: the base GLM-5.3-Flash is listed at about 6.7% on Harvey's agent benchmark on Vals AI's comparison pages. Per-criterion pass rates are a different, much higher number, so do not compare the two directly. Also note the benchmark itself has critics: a Hacker News commenter who inspected agent traces on Harvey's firm-knowledge subset reported finding problematic tasks and rubrics and linked an issue in Harvey's repository, so treat any single leaderboard number as a rough signal.
For how this kind of benchmark fits in the wider race, see our coverage of OpenAI's Astra for law and Gemini 4 Argon's launch benchmarks.
The base model underneath
GLM-Legal inherits its shape from GLM-5.3-Flash, the cheap-and-fast tier of Z.ai's GLM-5.3 family, released in late August 2026. We could not confirm the base model's parameter count or architecture from Z.ai's own pages, so we do not repeat the specs that circulated secondhand.
Vals AI's comparison pages list the base model at 83.93% on LegalBench, 45.19% on Legal Research Bench and about 6.7% on the Harvey agent benchmark. Those are the "before" numbers. The jump to 19.2% therefore suggests what domain mid-training and reinforcement learning in simulated legal environments can add on agent workflows. What we do not know is whether the specialization costs anything elsewhere: Z.ai has not published a full table, and legal research, which depends on how thoroughly a model searches case law, is the area to test directly if your work is research-heavy.
Why this matters beyond law
GLM-Legal is described as Z.ai's first specialized model for professional work, which suggests a product strategy: take a cheap general base, add a domain recipe made of corpora, simulated environments, a task harness and expert feedback, and sell access per vertical. Harvey and other legal-AI companies have been moving the same way, and the open question is whether vertical models from general labs can undercut specialist vendors on price while closing the quality gap. For buyers, the pressure is on cost per completed task, not cost per token, which is the number to track across vendors. For anyone building agents, the Harness idea is a reminder that the scaffold around the model, how it manages context, tools and feedback, is part of the product, not an afterthought.
The cost angle
The most interesting open question is price. Z.ai has published no GLM-Legal API price, and secondhand per-task cost comparisons against frontier models that circulated on social media could not be confirmed, so we do not repeat them. For reference, Vals AI lists the base GLM-5.3-Flash at roughly $0.57 per run on the Harvey benchmark, far below several frontier models. Z.ai's pitch, in line with its GLM-5.3 positioning, is frontier-adjacent quality at a fraction of the cost.
That is the same bet discussed in our look at why small and cheap models are catching up and in the startup open-weight cost savings story.
Is it open weight?
This is where the sources disagree.
- An aggregator headline calls GLM-Legal the top open-weight model on LegalBench and Harvey's benchmark.
- Unite.AI says access is via a consultation application and does not say whether weights are open.
- Hacker News commenters, including postalcoder, noted it does not look like an open model, and another asked whether sentiment toward Chinese models would shift if labs keep them proprietary.
Our read: GLM-Legal itself is in a gated beta, access is by application, and we found no weights release. If weights appear later, that would be an update worth a new post. For privacy-sensitive legal work, the distinction matters, as one commenter (consumer451) asked who processes and stores a client's private data when a model is reachable only through the vendor's service.
What developers and lawyers are saying
The Hacker News discussion mixes skepticism with curiosity:
- Hallucination risk. Commenter Natsu pointed out that the demo leans on tracked changes and human review but does not address fabricated cases or holdings, the failure that has repeatedly embarrassed lawyers in court.
- Benchmark skepticism. An attorney, otterley, said a roughly 20% pass rate does not impress them, since litigation is about novel situations rather than relitigating past ones. Another commenter, mathis-l, questioned the benchmark's quality.
- Jurisdiction. Commenters bb123 and arctic-true asked which legal systems the model is tuned for; the examples seemed oriented toward common-law countries. Z.ai's own example set included Southeast Asian privacy compliance.
- Domain models. Depurator wondered whether domain-specific models like this are a trend, with GLM-Legal as the first of many.
These are individual commenters' views, not verified facts.
How to run your own evaluation
A single leaderboard number says little about your matters, so build a small private test set before you commit. Pick twenty to thirty tasks your team already completed, such as a contract clause review, a privilege-neutral document summary and a compliance checklist, and keep the human-approved answers as the reference. Write the pass criteria down in advance, in the same all-or-nothing style Harvey LAB uses, and also record a looser per-criterion score so you can see where the model nearly succeeds.
Run the same set against GLM-Legal, its base model and one frontier model, and log three things for each: the pass rate, the cost per completed task, and the number of fabricated or unverifiable citations. That last count is the one lawyers care about most, because a single invented case can outweigh many correct answers. Repeat the run a few weeks later, since beta models change without notice and a result from launch week may not hold.
How to try it
- Submit the access form on Z.ai's site for API or collaboration access; the form asks how you plan to use the model.
- AutoClaw, Z.ai's desktop agent, is meant to be the quicker route, but Unite.AI reports Z.ai says GLM-Legal support there is coming "soon", so check availability before relying on it.
- Run your own evaluation on anonymized, non-privileged documents before any real matter.
What to do before you rely on it
Contract sheet with a magnifier on one bar, a checklist for verifying a legal AI model like GLM-Legal
- Treat outputs as drafts. Require a human to verify every citation and quotation.
- Test on your jurisdiction. A model tuned for common-law examples may behave differently elsewhere.
- Compare against a frontier baseline on your own tasks, with cost per task included.
- Check data-handling terms. A gated vendor API means your documents leave your environment.
- Keep an eye on independent evaluations from groups like Vals AI and Artificial Analysis. For the broader legal-AI picture, see our guide to using AI for legal help and contracts.
Related reading
- OpenAI Astra for law and legal research
- GLM-5.3 ties Kimi K3 on the AA Intelligence Index
- Gemini 4 Argon launch, benchmarks and pricing
- Small models have arrived
- Startups and open-weight model cost savings
- AI and law: legal help and contracts guide
- Hacker News discussion of GLM-Legal
Details are accurate as of October 12, 2026 and drawn from secondary summaries of Z.ai's announcement; check Z.ai's page for current terms.