- Z.ai announced GLM-Legal on October 12, 2026. It is built on GLM-5.3-Flash, an MIT-licensed open-weight model with 320 billion total and 18 billion active parameters, and aimed at contract review, evidence analysis and compliance research.
- Vals AI, which ran the model on October 12, ranks it 7th of 79 on Harvey's Legal Agent Benchmark with a 19.17% all-pass rate (±3.25), at $0.15 per task. The base GLM-5.3-Flash scored 6.67%.
- As of October 16, GLM-Legal is an application-only API beta. Z.ai has published no price or license for it, and no GLM-Legal repository appears on its Hugging Face page, even though Vals labels the weights 'Open.'
Contents
On October 12, 2026, Z.ai (formerly Zhipu AI) announced GLM-Legal, a model trained for legal work such as contract review, evidence analysis and compliance research. It is built on GLM-5.3-Flash, one of Z.ai's open-weight models, and an independent evaluator ranks it near the top of a new legal-agent benchmark at a fraction of rivals' cost. The catch: GLM-Legal itself isn't open, isn't priced and isn't generally available yet.
This story is based on Z.ai's launch blog post and launch post on X, Vals AI's model page and leaderboard, Harvey's benchmark announcement, and Z.ai's Hugging Face model card and pricing page, checked October 16, 2026. A Hacker News thread about the launch, posted October 12, had drawn 674 points by October 16.
The short version
- What it is: a legal-specialist model from Z.ai, built on GLM-5.3-Flash (320 billion total parameters, 18 billion active, MIT license).
- How it was trained, per Z.ai's launch post: more legal text in mid-training, then reinforcement learning "across tens of thousands of realistic legal environments," with feedback from legal professionals feeding later training cycles.
- Independent result: Vals AI ranks it 7th of 79 on Harvey's Legal Agent Benchmark (LAB), passing 19.17% of tasks in full, at $0.15 per task.
- Availability: an API beta you apply for. No public price, license or weights as of October 16. Z.ai says users will "soon" be able to use it in AutoClaw, its desktop agent app, and gives no date.
What is GLM-Legal?
Z.ai's launch post, titled "GLM-Legal: Advancing Agentic Legal Intelligence" and dated October 12, describes a two-step recipe on top of GLM-5.3-Flash. First, Z.ai says it "expanded the volume of high-quality legal corpora" during mid-training. Second, it trained the model with reinforcement learning inside an agent harness, and says "harness architecture and model training iterate in lockstep." Z.ai adds that in production, legal professionals evaluate outputs against criteria such as whether "all critical issues and edge cases" are addressed, and that it feeds those findings back into later training cycles.
Z.ai hasn't published its training data, so outsiders can't check how close its training environments were to public benchmark tasks. It invites law firms, legal-tech companies, in-house legal teams and independent practitioners to join the beta. Its launch post on X says the model "matches frontier models on mainstream legal benchmarks," and the blog claims it matches "frontier closed-source models on LegalBench and PLawBench." Those are vendor claims; the independent numbers are below.
The base model is better documented. Z.ai's GLM-5.3-Flash model card on Hugging Face lists an MIT license and 320 billion total parameters with 18 billion active. Z.ai gives no separate parameter count for GLM-Legal.
How did GLM-Legal score on legal benchmarks?
The most useful numbers come from Vals AI, an independent evaluation company, which published GLM-Legal results on October 12, 2026.
Harvey's Legal Agent Benchmark (LAB) was released by legal-AI company Harvey on May 6, 2026. It has more than 1,200 tasks across 24 practice areas and over 75,000 expert-written rubric criteria. Agents produce real files (documents, spreadsheets, slides) using file and shell tools, with internet access off. A task counts only if every criterion passes, so scores are low across the board. Vals grades the held-out test set with two LLM judges, GPT 5.5 and Claude Sonnet 4.6, and averages them.
| Model (Vals AI, Harvey LAB) | Task pass rate | Rank | Cost per task | Time per task |
|---|---|---|---|---|
| Muse Spark 1.2 | 25.42% | 1 | $2.09 | 24m16s |
| Gemini 4 Argon | 19.58% | 6 | $10.55 | 43m55s |
| GLM-Legal | 19.17% (±3.25) | 7 | $0.15 | 1h12m |
| Mistral Large 4 | 15.83% | 9 | $4.18 | 39m25s |
| GLM-5.3-Flash (base) | 6.67% | 31 | $0.57 | 52m17s |
Last verified: October 16, 2026 (Vals AI leaderboard, last updated October 12, 2026)
What the table does and doesn't show:
- The jump over the base model is large: from 6.67% to 19.17% on the same benchmark and harness.
- Ranks 6 and 7 are a statistical tie. Vals gives GLM-Legal a ±3.25-point margin, and Gemini 4 Argon is 0.41 points ahead.
- Cost comes with a footnote. Vals ran both GLM models at $0.075 per million input tokens and $0.25 per million output, a lower rate than Z.ai's current list price for GLM-5.3-Flash ($0.15 and $0.50). GLM-Legal has no list price, so its real cost per task is unknown.
- It's slow. At 1 hour 12 minutes per task, it is the slowest model in the table above.
On two other Vals benchmarks GLM-Legal is mid-pack: 86.23% on LegalBench (15th of 151) and 41.83% on Legal Research Bench (24th of 78), where MarkTechPost reports the base model scored higher (45.19%). Z.ai also says GLM-Legal matches frontier closed models on LegalBench and PLawBench, a Chinese-law benchmark; the PLawBench comparison is vendor-reported, and Z.ai's post credits Vals with the independent Harvey LAB and LegalBench results.
Can you download GLM-Legal? Is it open source?
Not as of October 16, 2026. The evidence:
- Z.ai's launch post and X post invite users to "Apply for API access" through a consultation form, under the heading "Join the GLM-Legal Beta."
- Z.ai's Hugging Face organization shows no GLM-Legal repository.
- Z.ai's pricing page doesn't list GLM-Legal.
- No license for GLM-Legal has been published.
There is one conflicting signal. Vals AI's model page labels the weights "Open" and calls it an "open-weight model from zAI." Z.ai hasn't said that, and nothing downloadable exists. Until Z.ai publishes weights and a license, treat GLM-Legal as a closed API model built on an open base.
GLM-5.3-Flash's MIT license lets you fine-tune and deploy it yourself; GLM-Legal's terms are unknown. Our guide to open-weight licenses explains why the license, not the word "open," decides what you can do.
What does GLM-Legal cost?
| Item | Price | Source |
|---|---|---|
| GLM-Legal API | Not published (beta by application) | Z.ai launch post |
| GLM-5.3-Flash input | $0.15 per 1M tokens | Z.ai pricing page |
| GLM-5.3-Flash cached input | $0.03 per 1M tokens | Z.ai pricing page |
| GLM-5.3-Flash output | $0.50 per 1M tokens | Z.ai pricing page |
Last verified: October 16, 2026
Why it matters
GLM-Legal is a clean example of a trend: a cheap open base model, post-trained for one profession, landing close to frontier models on that profession's benchmark. Our comparison of frontier APIs and open models covers why per-task cost, not per-token price, is the number that matters for agent work. And like Reflection AI's Beam, it shows how a company's own framing ("matches frontier models") can run ahead of what's actually shipped.
For legal teams, a top-ten LAB score is a reason to look, not to deploy. A 19% all-pass rate means about four in five tasks still had at least one failed criterion, and Harvey's grading gives no partial credit, so a near-miss and a bad answer look the same in the headline number. Lawyer review stays essential.
What's confirmed, reported and uncertain
- Confirmed (primary sources): the launch date, title and wording of Z.ai's blog and X posts; Vals AI's GLM-Legal results, dates and run settings; Harvey LAB's design; GLM-5.3-Flash's MIT license, size and list price; no GLM-Legal repository on Z.ai's Hugging Face page or price on its pricing page as of October 16.
- Vendor claims (Z.ai): the training recipe; "matches frontier models on mainstream legal benchmarks"; parity on LegalBench and PLawBench; AutoClaw support "soon."
- Uncertain: whether Z.ai will release GLM-Legal weights, and under what license; its API price; why Vals labels the weights "Open"; GLM-Legal's own parameter count.
About this storyBased on the sources linked below. Editorial standards
