Z.ai GLM-Legal Beats Claude Fable 5 and Kimi K3 on Harvey LAB
MarkTechPost Michal Sutter
Z.ai launched GLM-Legal, a law-focused model that scored 19.17% on Harvey LAB. It’s cheaper than Claude Fable 5, but you have to apply for access.
Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Z.ai has put out GLM-Legal, a legal model built on GLM-5.3-Flash and tuned for agentic work in law. The target use cases are familiar ones: contract review, evidence analysis, and compliance research. The headline number is the one Z.ai clearly wants people to see first. On Harvey’s Legal Agent Benchmark, GLM-Legal reached a 19.17% all-pass rate, up from 6.67% for the base model.
That’s not just a small bump. It lifts GLM-Legal into #7 out of 79 on Vals AI’s leaderboard, and Z.ai says the model lands ahead of Kimi K3, Claude Fable 5, and GPT-6 Astra on that benchmark. An independent evaluator reproduced the result, and Vals AI’s numbers line up with the launch chart. Still, the chart leaves out six models ranked above it, including Muse Spark 1.2.
The model is built on some serious hardware muscle underneath: 320B total parameters, 18B active, a 1M-token context window, and 128K max output. Z.ai says the base architecture mixes sparse and linear attention to keep long-context serving costs down. The training story also matters here. GLM-Legal went through mid-training on legal corpora, then post-training with reinforcement learning across tens of thousands of legal task environments.
The catch is access. GLM-Legal is API-only for now, and you have to apply. Z.ai hasn’t published a price, hasn’t released public weights, and says the model will come to its AutoClaw desktop agent app later, with no date attached. That makes the release feel less like an open launch and more like a controlled demo of what the company thinks legal AI should look like.
The comparison with the competition is mixed, which is usually how these things go once the launch gloss fades. GLM-Legal scored 86.23% on LegalBench, just behind Claude Fable 5 at 88.56%, and it fell to 41.83% on Legal Research Bench, below its own base model’s 45.19%. So yes, this is a real step up in one important benchmark. It is also a reminder that legal AI is still a patchwork of wins, gaps, and very selective bragging rights.
My take — AI-written commentary, not fact-checked reporting
This is the usual enterprise AI move: big benchmark headline, locked gate, no weights, no price, no real way to inspect the thing. Useful for lawyers, maybe, but not exactly a public good. Open models keep forcing the pace; closed ones keep charging for the velvet rope.
Read more about this at: MarkTechPost
Related stories
GLM-5.3’s Exploits, AI Models and Hardware Speed Up, DeepSeek’s New Agent Harness
The Batch ·
42