THE ESSENTIALS
  • Vals records a 19.17% Harvey benchmark pass rate, leaving substantial room for improvement.

Training for legal workflows

QbitAI's October 17 report examined GLM-Legal, Zhipu's model for contract review, evidence analysis and cross-border compliance research. Vals AI lists its release date as October 12. According to QbitAI, the model builds on GLM-5.3-Flash through additional legal training and reinforcement learning across tens of thousands of legal task environments.

What the scores establish

Vals records a 19.17% pass rate on Harvey's Legal Agent Benchmark and 86.23% accuracy on LegalBench. Harvey's benchmark requires every condition in a task to pass, so the first figure also shows that most tasks remain below that threshold. QbitAI reports that Zhipu is inviting law firms, legal technology companies and other practitioners to apply for its beta and provide feedback.

Sources & context

Go to the original material. Company claims remain attributed to their sources.

01
智谱新模型，懂法方面全球第一了 – 量子位 ↗The QbitAI report's visible header is 2026-10-17 20:13:45 Beijing time. Vals explicitly dates the model release to October 12; this distinction is retained in public prose. No overall global-first ranking or access/license claim is asserted. Retrieved publishedAt fields remain unchanged.
02
GLM Legal Benchmarks, Cost and Capabilities | Vals AI ↗The QbitAI report's visible header is 2026-10-17 20:13:45 Beijing time. Vals explicitly dates the model release to October 12; this distinction is retained in public prose. No overall global-first ranking or access/license claim is asserted. Retrieved publishedAt fields remain unchanged.
Last updated October 18, 2026.Spotted an issue? Let us know ↗