Release: a legal agent on a lightweight base

On October 13, Zhipu AI released GLM-Legal, a vertical LLM for the legal industry. According to Guandian and OSCHINA, the model is built on the lightweight GLM-5.3-Flash architecture, positioned as an agent model for legal workflows, and already matches closed-source peers on several legal-task benchmarks.

GLM-Legal — key facts

· Base: GLM-5.3-Flash (lightweight architecture)

· Positioning: agent model for legal workflows

· Performance: matches closed-source peers on legal benchmarks

· Cost: reportedly ~1/70 of comparable closed-source rivals

· Access: beta by application (per community reports)

Pass rate nearly tripled, cost cut to 1/70

BlockBeats, citing benchmark data, reported that GLM-Legal nearly tripled its legal-task pass rate over previous levels while cutting test costs to roughly one-seventieth of comparable closed-source rivals. Community tracking also shows a 19.2% pass rate on the Harvey LAB legal-agent benchmark. Unite.AI notes that GLM-Legal's goal is not "bigger chat" but chaining retrieval, drafting and review into executable agent workflows.

The "small base + specialized data + tiny cost" playbook is recreating the DeepSeek effect inside verticals: don't pile on parameters — make one job usable and cheap.

Why law is the first battleground for vertical models

Legal work is text-dense, error-intolerant and willing to pay — a combination already validated by Harvey and peers in the US. For law firms, corporate legal departments and government legal-affairs units, a domestically developed vertical model that can be deployed on-premise with auditable outputs fits data-compliance requirements far better than calling overseas general APIs. That helps explain why the first widely noted Chinese legal agent came from Zhipu, the lab best known for "lightweight and open".

The NineZenith view

NineZenith's own vertical-LLM practice rhymes with this: the base need not be the largest, but domain knowledge and compliance chains must be complete. The Tiandun safety stack's content-security and audit capabilities are precisely the precondition for deploying agents in heavily regulated fields like law — the vertical-agent contest is ultimately a product of "capability × cost × compliance".

(Compiled from public reports by Guandian, OSCHINA, BlockBeats, Unite.AI, etc.)