On October 12, Z.ai released GLM-Legal, an agentic legal intelligence model. It is not another general-purpose flagship: built on the lightweight GLM-5.3-Flash base and trained with legal corpora and tens of thousands of task environments, it is a "specialist" aimed at turning contract review, evidence analysis and compliance research into end-to-end agent deliverables.

The Score: 19.2% All-Pass, Cheapest in the Top Ten

In the official evaluation, GLM-Legal scored a 19.2% all-pass rate on Harvey LAB, a benchmark of real law-firm workflows across 24 practice areas — a 12.5 percentage-point improvement over its base model. Z.ai also reports parity with frontier closed-source models on LegalBench and the Chinese-language PLawBench.

Independent evaluator Vals puts it in concrete coordinates: 19.17% (±3.25) on Harvey LAB, ranking 7th of 79 models, at $0.15 per test — the cheapest model in the top ten. On LegalBench it scores 86.24%, ranking 14th of 150 and marking the best GLM-family result on that benchmark; on Legal Research Bench it reaches 41.83%, 24th of 78. Pricing is $0.075 per million input tokens and $0.25 per million output tokens, with a 1M context window and open weights.

Training: Mid-Training Foundations, RL Practice, Co-Evolving Harness

The most informative part of the release blog is not the scores but the idea of co-evolving the model with its agent harness. Training has three layers:

  • Mid-training expands high-quality legal corpora so the model learns statutes, interpretive frameworks and the fact-rule-conclusion deductive chain in dense domain context;
  • Post-training applies reinforcement learning across tens of thousands of realistic legal environments, teaching the model to plan analysis, advance workflows and produce deliverables under concrete constraints;
  • Production feedback has practitioners grade outputs against criteria like issue coverage, citation soundness and immediate usability, and the findings flow back into task environments, rubrics and harness design.

The blog stresses that the Harness — the runtime orchestration framework — is not an external tool but a first-class citizen iterating in lockstep with the model: edge cases found during training expose harness bottlenecks, while harness refinements feed realistic execution constraints back into the model.

In Practice: Contracts, Evidence and Compliance on AutoClaw

Alongside the release, Z.ai demonstrated three task types on AutoClaw, its desktop agent platform. The contract-review contrast is the most telling: where the base model suggests "consider shortening the notice period," GLM-Legal proposes executable edits — cutting the convenience-termination notice from 60 to 30 days, removing the cure period for IP, confidentiality or data-security breaches so the client can terminate immediately, and adding up to 90 days of transition assistance — written into the Word document with tracked changes. In evidence analysis it chains contradictory documents such as an external "60% complete" report against an internal ~41% snapshot, keeping facts and inferences distinct. In compliance research it turns privacy obligations across Singapore, Malaysia and Indonesia into a prioritized checklist with statutory references.

So What

Two observations. First, the recipe of "small base + domain mid-training + RL across tens of thousands of environments" is driving down the unit cost of vertical leaderboards — being the cheapest in the top ten shows scores and bills can be won at the same time. Second, a 19.2% all-pass rate means roughly four in five tasks still break somewhere along the chain, and legal agents remain far from taking cases independently. For law firms and in-house teams, though, the usable positioning today is clear: a tireless junior associate who takes over first-pass review and research. The specialist-model war has just begun.

(Sources: Z.ai blog, Vals evaluation)