Z.ai has officially launched GLM-Legal, its latest artificial intelligence offering meticulously developed for the legal sector. Built upon the GLM-5.3-Flash foundation, this specialized model targets demanding agentic legal work, including detailed contract review, comprehensive evidence analysis, and critical compliance research.
A notable achievement for GLM-Legal is its performance on Harvey's Legal Agent Benchmark (LAB), where it recorded an impressive 19.17% all-pass rate. This figure represents nearly a threefold improvement over its base model's 6.67%, underscoring the effectiveness of its specialized training. This benchmark success positions GLM-Legal as a formidable contender in the rapidly evolving legal AI landscape.
Understanding GLM-Legal's Architecture and Training
GLM-Legal leverages the GLM-5.3-Flash architecture, characterized by 320 billion total parameters with 18 billion active parameters within a Mixture-of-Experts (MoE) framework. This design supports a vast 1-million-token context window and a 128,000-token maximum output, making it suitable for handling extensive legal documents. The model also features native multimodal capabilities, processing text, images, video, and various file types to produce text outputs.
The development of GLM-Legal involved a two-stage training process. Initially, the model underwent extensive mid-training using expanded legal corpora to refine its ability to deduce conclusions from facts and rules. This was followed by post-training through reinforcement learning across tens of thousands of simulated legal task environments. The Z.ai team highlights a unique aspect where the agent harness co-evolved alongside the model, with insights from training edge cases leading to improvements in both the harness and the model's adaptability to realistic constraints.
Benchmark Performance: Harvey LAB and LegalBench
The Harvey LAB is an open-source benchmark featuring 1,250 tasks across 24 legal practice areas, graded by over 75,000 expert-written rubric criteria. An "all-pass" is awarded only if every criterion is met. On this challenging benchmark, GLM-Legal secured a 19.17% task pass rate, placing it seventh out of 79 evaluated models on Vals AI's leaderboard. An independent evaluation confirmed this significant lift over its base model.
Comparatively, GLM-Legal outperformed several established models in the legal AI space on Harvey LAB. It surpassed Kimi K3 (12.92%), Claude Fable 5 (11.25%), and GPT-6 Astra (5.42%). While these results are strong, it is worth noting that six other models currently rank higher, with Muse Spark 1.2 leading at 25.42%.
Regarding individual criteria, GLM-Legal achieved a 92.5% pass rate, aligning with the performance of other top-tier models which typically range from 90% to 95% on this metric.
On the LegalBench assessment, GLM-Legal registered an 86.23% score, demonstrating near parity with frontier models. This places it closely behind Claude Fable 5, which leads at 88.56%, and slightly ahead of Kimi K3's 86.21%.
Cost-Effectiveness and Access
One of GLM-Legal's most compelling attributes is its remarkable cost-efficiency. Vals AI reported a cost of approximately $0.15 per Harvey LAB test for GLM-Legal, which stands in stark contrast to the estimated $19.23 for Claude Fable 5 or $26.16 for GPT-6 Astra. This makes GLM-Legal significantly more economical for similar high-stakes legal tasks.
Currently, access to GLM-Legal's API is restricted to an application-only basis for law firms, legal-tech companies, and in-house counsel, with pricing not yet publicly disclosed. While the base model, GLM-5.3-Flash, is available for broader use at $0.15 per million input tokens and $0.50 per million output tokens, GLM-Legal's specific public release details, including weights and Hugging Face availability, remain unannounced. Z.ai has indicated that GLM-Legal will eventually be integrated into AutoClaw, its desktop agent application, though a release date has not been provided.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
