Z.ai GLM-Legal Outperforms Claude Fable 5 and Kimi K3 on Harvey LAB
Z.ai has released GLM-Legal, a specialized legal AI model built on GLM-5.3-Flash and fine-tuned for agentic legal work. It targets contract review, evidence analysis, and compliance research. The headline result: a 19.2% all-pass rate on Harvey's Legal Agent Benchmark (LAB) โ nearly 3x the 6.7% scored by its base model, and enough to edge past frontier competitors Claude Fable 5 and Kimi K3.
The catch for developers is access: the API is currently application-only.
TL;DR
- Size: Built on GLM-5.3-Flash, a 320B-total, 18B-active Mixture-of-Experts (MoE) model with a 1M-token context window and 128K max output. Z.ai does not list a separate parameter count for GLM-Legal.
- Runs on: Z.ai's hosted API only, by application. No public weights. The base model can be self-hosted on multi-GPU servers via vLLM or SGLang with FP8 weights.
- Performance: Top-10 on Harvey LAB at a fraction of frontier model costs. Within 2.3 points of the LegalBench leader.
- Best: 19.17% Harvey LAB task pass rate and #7 of 79 on Vals AI, at $0.15 per test.
- Worst: 41.83% on Vals AI's Legal Research Bench, below its own base model (45.19%).
What Is GLM-Legal?
GLM-Legal is a Z.ai model derived from GLM-5.3-Flash, a 320B-total, 18B-active MoE foundation with a 1M-token context window and 128K maximum output. Rather than scale parameters further, Z.ai focused on post-training for legal reasoning, document drafting, and multi-step agent tasks. The model is optimized for workflows that require parsing long contracts, cross-referencing evidence, and researching compliance obligations.
Harvey LAB Performance
Harvey's Legal Agent Benchmark evaluates models on realistic legal tasks with pass/fail scoring. GLM-Legal achieved a 19.2% all-pass rate, a substantial jump from the 6.7% posted by its GLM-5.3-Flash base. That gain places it ahead of Claude Fable 5 and Kimi K3 on the same benchmark.
In plain terms, GLM-Legal is not the overall leader โ but it delivers top-10 results at a much lower cost point, making it attractive for high-volume legal review pipelines where frontier pricing is prohibitive.
Vals AI Results
Independent evaluations on Vals AI show mixed but promising outcomes:
- Rank: #7 of 79 models tested
- Cost: $0.15 per test
- Legal Research Bench: 41.83% pass rate โ below the 45.19% scored by the GLM-5.3-Flash base
The Legal Research Bench result suggests the legal fine-tuning trades some broad research capability for stronger performance on task-oriented agentic work such as contract review and evidence analysis.
Key Specs at a Glance
| Attribute | Detail |
|---|---|
| Base model | GLM-5.3-Flash |
| Architecture | 320B-total, 18B-active MoE |
| Context window | 1M tokens |
| Max output | 128K tokens |
| Harvey LAB all-pass | 19.2% |
| Vals AI ranking | #7 of 79 |
| Cost per test | $0.15 |
| Availability | Hosted API, application-only |
| Public weights | No |
Developer Access and Deployment
GLM-Legal is available only through Z.ai's hosted API, and access requires an application. There are no public weights for the legal variant. Developers who want self-hosting can deploy the GLM-5.3-Flash base model on multi-GPU servers using vLLM or SGLang with FP8 quantization โ but that path does not include the legal fine-tuning that drives the benchmark gains.
Why It Matters
Legal AI has become one of the most competitive enterprise categories heading into 2026, with specialized benchmarks like Harvey LAB and Vals AI separating general-purpose frontier models from domain-tuned alternatives. GLM-Legal's results show that targeted post-training on a strong MoE base can close much of the gap with frontier systems โ at roughly a fraction of the inference cost. For teams evaluating legal automation, the trade-off is clear: best-in-class research breadth versus cost-efficient, task-focused agentic performance.
via MarkTechPost