# GLM-Legal

> Source: https://aiwiki.ai/wiki/glm_legal
> Updated: 2026-10-15
> Categories: AI Models, Chinese AI, Large Language Models, Legal AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "GLM-Legal." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/glm_legal
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**GLM-Legal** is a [large language model](https://aiwiki.ai/wiki/large_language_model) for legal work developed by [Z.ai](https://aiwiki.ai/wiki/z_ai), the international brand of [Zhipu AI](https://aiwiki.ai/wiki/zhipu_ai). Z.ai announced it on October 12, 2026, describing it as a model built on [GLM-5.3-Flash](https://aiwiki.ai/wiki/glm_5_3_flash) and designed for tasks such as contract review, evidence analysis and compliance research.[1][2] The company says GLM-Legal matches frontier models on mainstream legal benchmarks. At launch it was offered as a beta: organizations could apply for API access through a consultation form, and Z.ai said the model would soon be available in AutoClaw, its desktop agent product.[1][3]

[Vals AI](https://aiwiki.ai/wiki/vals_ai), an independent benchmarking company, evaluated GLM-Legal on the day of the announcement. It measured a 19.17% task pass rate on Harvey's Legal Agent Benchmark and 86.23% on [LegalBench](https://aiwiki.ai/wiki/legalbench), at a listed price of US$0.075 per million input tokens and US$0.25 per million output tokens.[5][6][7]

## Key facts

| Field | Detail |
| --- | --- |
| Developer | Z.ai (Zhipu AI)[1] |
| Announced | October 12, 2026[1][2] |
| Base model | GLM-5.3-Flash[1] |
| Intended uses | Contract review, evidence analysis, compliance research and other legal workflows[1][2] |
| Access at launch | Beta; API access by application through Z.ai's consultation form[1][3] |
| Context window (Vals AI listing) | 1M tokens[5] |
| Maximum output (Vals AI listing) | 128k tokens[5] |
| Input (Vals AI listing) | Text only[5] |
| Price (Vals AI) | US$0.075 per million input tokens, US$0.25 per million output tokens[5][6] |

## Background

Z.ai frames GLM-Legal as part of a broader shift in language models from single-turn code generation toward what it calls "Professional Cowork": work in which a model has to digest fragmented source files under concrete objectives and delivery constraints. In its announcement the company argues that legal practice concentrates these demands, because it requires an unbroken chain between facts, statutory doctrine, analysis and client-ready work product, and it treats legal work as a stress test for training models for high-stakes professional settings.[1]

The base model, GLM-5.3-Flash, was released by Z.ai in August 2026. Its model card describes it as a natively multimodal model with 320 billion total parameters and 18 billion active parameters, released under the MIT License.[16] Z.ai's GLM-Legal announcement does not give a separate parameter count for GLM-Legal; it says only that the model runs on "a significantly leaner parameter footprint" than the frontier closed-source models it is compared with.[1]

## Training

Z.ai describes three stages in its training approach.[1]

- **Mid-training.** Z.ai says it expanded the volume of high-quality legal corpora so the model could learn legal concepts, statutory rules and interpretive frameworks, with the aim of strengthening the deductive links between facts, rules and conclusions.
- **Post-training.** The company says it used [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) across "tens of thousands of realistic legal environments", teaching the model to plan an analysis, move a workflow forward and produce deliverables under concrete constraints. Z.ai describes the agent harness (the runtime framework that manages task context, connects tools and routes execution feedback) as being developed together with the model, so that problems found during training informed changes to the harness and vice versa.
- **Production feedback.** Z.ai says practitioner feedback is turned into training signals, using criteria such as whether all critical issues are addressed, whether citations are legally sound, whether the output follows the matter brief and whether it can be used without manual rework.

The announcement does not disclose the size or sources of the legal training data.[1] The legal-technology publication Artificial Lawyer, quoting Z.ai, described the model as based on GLM-5.3-Flash "with additional legal-domain training and reinforcement learning", and said it is multilingual.[11]

## Benchmark results

### Z.ai's reported results

Z.ai's launch post compares GLM-Legal with its base model and three other models on three legal benchmarks. It states that the Harvey's Legal Agent Benchmark and LegalBench results were independently evaluated by Vals AI.[1]

| Model | Harvey's Legal Agent Benchmark | LegalBench | PLawBench |
| --- | --- | --- | --- |
| GLM-Legal | 19.2 | 86.2 | 77.3 |
| GLM-5.3-Flash | 6.7 | 83.9 | 73.8 |
| [Kimi K3](https://aiwiki.ai/wiki/kimi_k3) | 12.9 | 86.2 | 76.2 |
| [Claude Fable 5](https://aiwiki.ai/wiki/claude_fable_5) | 11.3 | 88.6 | 77.7 |
| [GPT-6 Astra](https://aiwiki.ai/wiki/gpt_6_astra) | 5.4 | 87.1 | 69.1 |

*Scores as shown in Z.ai's "Legal Benchmark Evaluation" chart.[1]*

The three benchmarks measure different things:

- **Harvey's Legal Agent Benchmark (LAB)** is an open-source benchmark from the legal AI company [Harvey](https://aiwiki.ai/wiki/harvey_ai) that tests agents on legal assignments in realistic environments. Its repository contains tasks with agent instructions, documents and rubrics, together with an execution harness, and covers 24 legal practice areas plus contracting.[13] Under Harvey's all-pass grading, a task counts as passed only if every rubric criterion passes.[13]
- **LegalBench** is a collaboratively built benchmark of legal reasoning tasks introduced by Neel Guha and co-authors in 2023.[15]
- **PLawBench** is a rubric-based benchmark of legal practice published in January 2026. It contains 850 questions across 13 practical scenarios in three task categories (public legal consultation, practical case analysis and legal document generation), with about 12,500 expert-designed rubric items. Its authors state that it is grounded in the Chinese legal system and that all scenarios are written in Chinese.[14] Z.ai's post attributes independent evaluation by Vals AI only to the Harvey's Legal Agent Benchmark and LegalBench results.[1]

On LegalBench and PLawBench, GLM-Legal's reported scores are close to but below Claude Fable 5's; Z.ai describes this as "parity" with leading frontier systems.[1] The comparison set in the chart is Z.ai's choice, and it does not include every model that scored higher on these benchmarks (see below).

### Vals AI evaluation

Vals AI ran GLM-Legal on three of its legal benchmarks. Its model page lists Z.ai as the default provider and default settings of temperature 1, top-p 0.95, a 128,000-token output limit and maximum reasoning effort, noting that some benchmarks may use different providers or parameters.[5]

| Benchmark (Vals AI) | GLM-Legal score | Rank shown on Vals AI | Notes |
| --- | --- | --- | --- |
| Harvey's Legal Agent Benchmark | 19.17% task pass rate | 7th of 79 | 92.5% of individual criteria met; about US$0.15 per test; listed duration 1 h 12 min[5][6] |
| LegalBench | 86.23% | 15th of 151 | Latency 6.00 s[5][7] |
| Legal Research Bench | 41.83% all-pass | 24th of 78 | Tied in score with Claude Sonnet 5 (23rd); about US$0.05 per test; listed duration 39 min 49 s[5][8] |

Rankings on Vals AI's leaderboards change as models are added. In its October 12 update note, Vals AI listed GLM-Legal as 14th of 150 on LegalBench (86.24%), and in an X post that day it described the model as 6th overall on Harvey's benchmark. The leaderboards consulted on October 15, 2026 showed the positions in the table above.[5][6][7][10]

On Harvey's benchmark, Vals AI's leaderboard placed six models above GLM-Legal, led by Muse Spark 1.2 at 25.42%; the sixth, [Gemini 4 Argon](https://aiwiki.ai/wiki/gemini_4_argon), scored 19.58% at about US$10.55 per test. Vals AI called GLM-Legal the cheapest model in the top 10 on that benchmark, and estimated that it cost roughly 70 times less per task than Gemini 4 Argon.[5][6][10] Compared with Z.ai's own models, Vals AI reported that GLM-Legal's 19.2% was about three times GLM-5.3-Flash's 6.7% and more than double [GLM-5.3](https://aiwiki.ai/wiki/glm_5_3)'s 8.3%, and that the share of individual criteria met rose from 82.2% for GLM-5.3-Flash to 92.5%.[10] On LegalBench, GLM-Legal (86.23%) sat one place above Kimi K3 (86.21%); the top score on that leaderboard was Claude Fable 5's 88.56%.[7]

In an X post on October 12, Vals AI described GLM-Legal as "the new #1 open-weight model on Harvey's Legal Agent Benchmark and LegalBench".[9]

## Availability and weights

Z.ai's launch post invited law firms, legal technology companies, in-house legal teams and independent practitioners to join a GLM-Legal beta, with API access available by application through a form on Z.ai's site.[1][3] The post does not state a license for the model or mention a weights download.[1]

Sources differ on the model's openness. Vals AI's model page lists GLM-Legal's weights as "Open" and its X post called it an open-weight model.[5][9] Artificial Lawyer reported on October 13 that GLM-Legal is understood not to be intended, "at least not yet", as an open-weights model, unlike GLM-5.3-Flash.[11] As of October 15, 2026, Z.ai's Hugging Face organization (zai-org) listed GLM-5.3-Flash but no GLM-Legal repository.[16][17]

## AutoClaw integration

AutoClaw is a general AI agent developed by Z.ai that can work across files, browsers and tools; its site offers downloads for macOS, Windows and Linux.[4] Z.ai's announcement describes it as a desktop-native agent platform and says users would "soon" be able to use GLM-Legal there.[1] The post demonstrates three scenarios, each comparing GLM-Legal's output with GLM-5.3-Flash's on the same task:[1]

| Scenario | Task given in Z.ai's example |
| --- | --- |
| Contract review | Review a software services contract for a client switching provider, focusing on termination and handover; AutoClaw applies the suggested edits to the Word document as tracked changes and comments |
| Evidence analysis | Analyze 40 documents from a software project dispute to find contradictions in the provider's progress reports |
| Compliance research | Plan privacy compliance for a website launching in Singapore, Malaysia and Indonesia, with tracking, email marketing and overseas cloud storage |

These examples are Z.ai's own demonstrations.

## Reception

Artificial Lawyer's coverage focused on the fact that a general-purpose model developer had trained and launched its own legal-specific model. It wrote that this was "a LLM maker launching their own legal-specific model", not an add-in or an invitation for legal technology companies to build on its platform, and noted that lawyers would want more information about where their data goes and that costs were not yet clear.[11] The Japanese technology site GIGAZINE reported the launch on October 13, noting the beta status, the planned AutoClaw availability and the API application process.[12]

## See also

- [Z.ai](https://aiwiki.ai/wiki/z_ai)
- [GLM-5.3-Flash](https://aiwiki.ai/wiki/glm_5_3_flash)
- [LegalBench](https://aiwiki.ai/wiki/legalbench)
- [Vals AI](https://aiwiki.ai/wiki/vals_ai)
- [Harvey](https://aiwiki.ai/wiki/harvey_ai)
- [AI in law](https://aiwiki.ai/wiki/ai_in_law)

## References

1. Z.ai, "GLM-Legal: Advancing Agentic Legal Intelligence," Z.ai blog, October 12, 2026. https://z.ai/blog/glm-legal
2. Z.ai (@Zai_org), "Introducing GLM-Legal," X, October 12, 2026. https://x.com/Zai_org/status/2109678174468534511
3. Z.ai, "Contact Us for Your Legal AI Needs" (GLM-Legal API access application). Accessed October 15, 2026. https://z.ai/consultation/legal
4. Z.ai, "AutoClaw: An AI agent for Work." Accessed October 15, 2026. https://autoclaw.z.ai
5. Vals AI, "GLM Legal Benchmarks, Cost and Capabilities," including the update "zAI's GLM Legal evaluated on our legal benchmarks" (October 12, 2026). Accessed October 15, 2026. https://www.vals.ai/models/zai_glm-legal
6. Vals AI, "Harvey's Legal Agent Benchmark Leaderboard," updated October 12, 2026. Accessed October 15, 2026. https://www.vals.ai/benchmarks/hlab
7. Vals AI, "LegalBench" leaderboard, updated October 13, 2026. Accessed October 15, 2026. https://www.vals.ai/benchmarks/legal_bench
8. Vals AI, "Legal Research Bench" leaderboard, updated October 12, 2026. Accessed October 15, 2026. https://www.vals.ai/benchmarks/legal_research
9. Vals AI (@ValsAI), post on GLM Legal, X, October 12, 2026. https://x.com/ValsAI/status/2109678919108690380
10. Vals AI (@ValsAI), post on GLM Legal's Harvey's Legal Agent Benchmark results, X, October 12, 2026. https://x.com/ValsAI/status/2109678921017168061
11. Artificial Lawyer, "Chinese Company Launches Legal-Specific LLM," October 13, 2026. https://artificiallawyer.com/2026/10/13/chinese-company-launches-legal-specific-llm/
12. GIGAZINE, "GLM-Legal, an AI specializing in legal tasks, has been released. This Chinese-made model, based on GLM-5.3-Flash, boasts performance comparable to GPT-6 Astra," October 13, 2026. https://gigazine.net/gsc_news/en/20261013-glm-legal/
13. Harvey, "harvey-labs: Legal Agent Benchmark (LAB)," GitHub repository. Accessed October 15, 2026. https://github.com/harveyai/harvey-labs
14. Shi, Yuzhen, et al., "PLawBench: A Rubric-Based Benchmark for Evaluating LLMs in Real-World Legal Practice," arXiv:2601.16669, January 2026. https://arxiv.org/abs/2601.16669
15. Guha, Neel, et al., "LegalBench: A Collaboratively Built Benchmark for Measuring Legal Reasoning in Large Language Models," arXiv:2308.11462, 2023. https://arxiv.org/abs/2308.11462
16. Z.ai, "GLM-5.3-Flash" model card, Hugging Face. Accessed October 15, 2026. https://huggingface.co/zai-org/GLM-5.3-Flash
17. Hugging Face, models published by zai-org. Accessed October 15, 2026. https://huggingface.co/zai-org