GLM-Legal Scores 19.2% on Harvey LAB

GLM-Legal’s domain specialization is more compelling than the “beats every frontier model” framing, but the benchmark result still shows how much targeted post-training can improve agent reliability.

• –The 19.2% all-pass score measures complete legal deliverables, not isolated question answering.
• –Vals ranks GLM-Legal sixth of 77 models, so the claim that it leads Claude Fable 5 and GPT-6 Astra should be treated cautiously.
• –Harvey’s benchmark evaluates long-horizon workflows across documents, tools, rubrics, and multiple legal practice areas.
• –Z.ai says the model combines legal mid-training with reinforcement learning in realistic task environments.
• –Open weights could make specialized legal agents cheaper to audit, fine-tune, and deploy privately.