Zhipu Open-Sources GLM-Legal: Harvey LAB #1 Open-Weight, How to Choose Legal AI?

TL;DR

Zhipu open-sourced GLM-Legal on October 17, 2026 — a legal-specialized model based on GLM-5.3-Flash, trained on legal corpora and reinforced with tens of thousands of legal tasks. It ranks #1 open-weight on both Harvey LAB (19.2% full pass rate) and LegalBench (86.23% accuracy) in independent third-party evaluation by Vals AI. More critically, the test cost is approximately 1/70th of Gemini 4 Argon at similar performance, trading about 30 minutes slower per task. Legal scenarios that prioritize reliability over speed suggest "cheap" and "fast" aren't the only competitive dimensions.

Model Specifications

Base & Training

Benchmark Results

Key insight: 19.2% vs the base model's 6.7% is a 12.5-point improvement — legal-specific training brought substantial gains.

Cost Comparison

A 70x cost difference means: if your legal tasks don't need results in under 30 minutes, GLM-Legal can reduce legal AI operating costs to 1.4% of the original.

What is Harvey LAB

Harvey LAB is a legal AI evaluation benchmark published by Vals AI. Its test design is particularly demanding: each task carries a chain of scoring conditions — miss one, and the entire task scores zero.

Test Design

• Over 1,200 simulated law firm work tasks
• Covers 24 practice areas: M&A, intellectual property, banking & finance, etc.
• Full pass = all scoring conditions met; otherwise zero Independently evaluated by third-party organization Vals AI, not self-reported by Zhipu

What 19.2% Means

A 19.2% full pass rate means approximately 1 in 5 tasks can be delivered "flawless and ready to use." This sounds low, but given Harvey LAB's scoring standard — miss one condition and the entire task scores zero — this is the highest level among open-weight models. More importantly, it's 12.5 percentage points higher than the base GLM-5.3-Flash, demonstrating that legal-specific training brought substantial improvement.

What It Can Do: Three Practical Scenarios

1. Contract Review

The most compelling scenario in official demos. GLM-Legal doesn't just say "suggest shortening the notice period" — it provides a revised draft with legal reasoning: shorten the termination notice from 60 days to 30 days, and add up to 90 days of transition support. This output style — specific revision suggestions + legal basis — is a key step from "prompting tool" to "drafting assistant" for legal AI.

2. Evidence Analysis

From 40 scattered documents, GLM-Legal connected a contradiction — client-facing report claimed 60% completion while internal snapshots showed only ~41% — into an arguable chain of evidence. This capability — finding contradictions across large document sets and constructing arguments — is a core paralegal function.

3. Cross-Border Compliance

For a website launching in Singapore, Malaysia, and Indonesia, GLM-Legal directly outputs a compliance action list sorted by P0 (must-complete before launch) priority. Chinese practice, English tasks, Southeast Asian cross-border — it handles all three.

What It Can't Do

The flip side of 19.2% is: there's still distance from "fully reliable." Real-world practice output still requires professional attorney oversight — the model is an assistant, not a replacement.

Zhipu's own actions reflect this positioning:

• Inviting law firms, legal tech companies, corporate legal teams, and independent practitioners to apply for GLM-Legal Beta
• Coming integration into desktop-native agent platform AutoClaw
• For ordinary users: let it be the first reader of contracts; the final signature stays with a human

How to Use: Developer Action Guide

1. If you're a law firm or legal tech company

Apply for GLM-Legal Beta and test it in your actual case workflows. Focus on contract review and evidence analysis — these offer the highest ROI as entry points.

2. If you're a corporate legal team

Use GLM-Legal as the "first reader" for contracts. Let it pre-screen contract terms, flag risk points and suggested revisions, then have legal staff review. This can improve contract review efficiency by an order of magnitude.

3. If you're an independent developer

GLM-Legal has open weights and can be deployed locally. Combined with cross-border compliance scenarios, you can build automated compliance checking tools — input a list of target markets, output a prioritized compliance action list.

4. Cost optimization strategy

If your legal tasks don't need results within 30 minutes, using GLM-Legal instead of Gemini 4 Argon can cut costs to 1/70th. For scenarios requiring batch contract review or evidence organization, this cost difference is decisive.

5. Managing expectations correctly

A 19.2% full pass rate means 80.8% of tasks still require human intervention. Positioning GLM-Legal as an "efficiency-boosting assistant" rather than a "lawyer replacement" is the most pragmatic approach at the current stage. AI entering law doesn't first change whether lawyers lose jobs — it changes the era where you'd sign contracts you couldn't understand.

Industry Significance

Specialization is the trend

GLM-Legal represents an important direction in AI model development: specialization. The ceiling for general-purpose LLMs in legal scenarios isn't about "knowing more" — it's about "understanding deeper." The two-step training method (legal corpora mid-training + legal task RL) transforms the model from "a general model that knows law" to "a legal-specialized model."

Cost curves replace parameter curves

Published the same day as DeepSeek V4.1-Flash-1017, both paths differ but point in the same direction: competition is shifting from "who's smarter" to "who can keep agents working without going bankrupt." GLM-Legal uses specialization to cut legal task costs to 1/70th of Gemini 4 Argon; DeepSeek uses Harness optimization to compress agent task runtime and output tokens.

The value of third-party evaluation

These results were independently evaluated by Vals AI, not self-reported by Zhipu. This is crucial — in an industry awash with self-reported data, independent third-party evaluation carries far more credibility. The numbers 19.2% and 86.23% therefore have genuine reference value.