Z.ai 发布 GLM-Legal：Harvey LAB 通过率 19.17%，超越 Claude Fable 5 与 Kimi K3

Z.ai just released GLM-Legal. GLM-Legal is a legal AI model from Z.ai, built on GLM-5.3-Flash and trained for agentic legal work. It targets contract review, evidence analysis and compliance research. The key result is a 19.2% all-pass rate on Harvey’s Legal Agent Benchmark (LAB). That is nearly 3x its base model’s 6.7%. For developers, the catch is access: the API is application-only for now.

TL;DR

Size: Built on GLM-5.3-Flash, a 320B-total, 18B-active MoE with a 1M-token context and 128K max output. Z.ai lists no separate GLM-Legal parameter count; Runs on: Z.ai’s hosted API only, by application. No public weights. The base model self-hosts on multi-GPU servers via vLLM or SGLang, with FP8 weights; Performance: Top-10 on Harvey LAB at a fraction of frontier cost. Within 2.3 points of the LegalBench leader; Best: 19.17% Harvey LAB task pass rate, #7 of 79 on Vals AI, at $0.15 per test; Worst: 41.83% on Vals AI’s Legal Research Bench, below its own base model (45.19%); Bottom line:

Best: frontier-class legal agent scores at about 1/128 the cost of Claude Fable 5; Worst: gated beta with no price, no weights and slow runs (1h12m per task)

What is GLM-Legal?

GLM-Legal is a domain-specialized version of GLM-5.3-Flash for law firms, in-house counsel and compliance teams. The base is Z.ai’s first natively multimodal GLM-5 model. It has 320B total parameters with 18B active. It mixes sparse and linear attention to cut long-context serving cost. It reads text, images, video and files, and outputs text.

How was GLM-Legal trained?

Training ran in 2 stages:

Mid-training on expanded legal corpora, to strengthen reasoning from facts to rules to conclusions; Post-training with reinforcement learning across tens of thousands of legal task environments

The Z.ai team also describes the agent harness co-evolving with the model during post-training. Edge cases in training exposed harness bottlenecks. Harness fixes then exposed the model to more realistic constraints.

How does GLM-Legal score on Harvey’s Legal Agent Benchmark?

Harvey LAB is an open-source benchmark of 1,250 tasks across 24 practice areas. Over 75,000 expert-written rubric criteria grade the work. A task passes only if every criterion passes. Harvey calls this all-pass grading.

Vals AI runs the held-out set with 2 LLM judges and internet disabled. Agents get file tools plus docx, pptx and xlsx skills.

On that leaderboard, GLM-Legal scores 19.17% and ranks #7 of 79. Its base, GLM-5.3-Flash, scores 6.67%. The 12.5-point lift is the core claim, and an independent evaluator reproduced it.

Z.ai’s launch chart shows GLM-Legal ahead of Kimi K3 (12.92%), Claude Fable 5 (11.25%) and GPT-6 Astra (5.42%). Those figures match Vals AI. The chart omits 6 models ranked higher, led by Muse Spark 1.2 at 25.42%.

GLM-Legal also passes 92.5% of individual criteria. Top models sit near 90% to 95% on that measure.

Does GLM-Legal match frontier models on LegalBench?

Roughly. Z.ai claims parity on LegalBench and the Chinese-focused PLawBench without publishing numbers. Vals AI scores GLM-Legal at 86.23% on LegalBench. Claude Fable 5 leads at 88.56%. Kimi K3 scores 86.21%.

How does GLM-Legal compare with Kimi K3, Claude Fable 5 and GPT-6 Astra?

All benchmark scores, costs and durations are from Vals AI’s runs as of October 12, 2026. Vals ran both GLM models at $0.075 / $0.25 per 1M tokens.

How can developers access GLM-Legal?

Law firms, legal-tech companies and in-house teams can apply for API access. Z.ai has published no price. No GLM-Legal repository exists on Hugging Face as of October 13, 2026. Z.ai says GLM-Legal will come to AutoClaw, its desktop agent app, but gave no date.

The base model is available today at $0.15 / $0.50 per 1M tokens. Teams can prototype pipelines on it, then swap the model ID later.

Key Takeaways

GLM-Legal scores 19.17% on Harvey LAB, up from 6.67% for its base; It ranks #7 of 79 on Vals AI, behind 6 Meta and Google models; It costs $0.15 per Harvey LAB test, versus $19.23 for Claude Fable 5; Legal Research Bench dropped 3.36 points versus GLM-5.3-Flash; Access is application-only, with no published price or weights

Michal Sutter

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.

Michal SutterMeta Launches Muse Wonder and Muse Wonder Full: Benchmarks, Pricing and How to Use It; Michal SutterTop 12 AI Live Translation Tools in 2026: Latency, Languages, Pricing; Michal SutterWhat is Decision 3.0? vLLM Semantic Router’s New Open Decision Models; Michal SutterMeet the Underdog Saluki 27B: A 2-bit Qwen3.8-27B That Beats the Original at Tool Calling