Physix Frontier · News Briefing Card (QbitAI · Oct 17, 2026)

Zhipu Releases GLM-Legal Open-Source Legal Model

KEY FACTS

Zhipu launched GLM-Legal, an open-source model purpose-built for legal work that can review contracts, connect evidence, and conduct cross-border compliance research.
On the Vals AI legal benchmark, GLM-Legal topped all open-weight models with a 19.2% full-pass rate and 86.23% accuracy.
The first version of Harvey LAB includes more than 1,200 simulated law-firm tasks spanning 24 legal practice areas, including M&A and intellectual property.
GLM-Legal was trained on the GLM-5.3-Flash base, with mid-training on legal corpora and reinforcement learning across tens of thousands of tasks.
Zhipu has opened applications for the GLM-Legal Beta and plans to integrate its capabilities into AutoClaw to drive real-world adoption.

KEY DATA

19.2%Task full-pass rate

86.23%Accuracy

over 1,200Harvey LAB simulated tasks

24Legal practice areas covered

PHYSIX OBSERVATION

Legal AI is moving from chat Q&A toward deliverable practice work, and Zhipu is cutting into law-firm scenarios with open source plus vertical training. The 19.2% full-pass rate shows the model is still early, but its cost is only about one-seventieth that of competitors, a clear value-for-money signal. For the industry, a general base plus specialized post-training may become the mainstream path for vertical models, and legal professionals should join trials and give feedback early rather than passively adapt once the tools have set.

Source: QbitAI report