GLM-Legal: Z.ai's Agentic Legal Intelligence — 19.2% on Harvey LAB, Frontier Parity on a Lightweight MoE Foundation
On October 12, 2026, Z.ai (formerly Zhipu AI) released GLM-Legal, a legal-domain agentic model built on the lightweight GLM-5.3-Flash foundation (320B total, 18B active parameters). It scores 19.2% all-pass on Harvey LAB (ranking #7 of 79), 86.2 on LegalBench, and 77.3 on PLawBench — matching frontier closed-source models on mainstream legal benchmarks while running at $0.075/$0.25 per million tokens. The model demonstrates that specialized mid-training and post-training can close the gap between lean open-weight models and expensive frontier systems on high-stakes professional tasks.
GLM-Legal: Z.ai's Agentic Legal Intelligence — 19.2% on Harvey LAB, Frontier Parity on a Lightweight MoE Foundation
Executive Summary
On October 12, 2026, Z.ai (formerly Zhipu AI) released GLM-Legal, a legal-domain agentic model that demonstrates a compelling approach to specialized AI: build on a lean, efficient foundation and close the performance gap through targeted mid-training and post-training rather than raw parameter scale. Built on GLM-5.3-Flash — a 320-billion-parameter Mixture-of-Experts model with only 18 billion active parameters — GLM-Legal achieves 19.2% all-pass on the Harvey Legal Agent Benchmark (ranking #7 of 79 models), 86.2 on LegalBench, and 77.3 on PLawBench, matching or approaching frontier closed-source models across all three axes.
What makes GLM-Legal notable is not just its performance but its efficiency: at $0.075 per million input tokens and $0.25 per million output tokens (inheriting GLM-5.3-Flash's pricing), it delivers frontier-tier legal reasoning at a fraction of the cost of models like Gemini 4 Argon ($2/$10) or Meta Muse Wonder ($4/$20). This represents a strategic pivot in the legal AI space — rather than training ever-larger general-purpose models and hoping they handle legal work well enough, Z.ai demonstrated that a focused training pipeline on a lean foundation can compete with the best.
The model is currently in API beta (application required via z.ai/consultation/legal) and will soon be available in AutoClaw, Z.ai's desktop-native agent platform. It covers contract review, evidence analysis, compliance research, and cross-border legal workflows — with particular strength in Chinese legal practice (PLawBench) while maintaining strong English-language performance.
1. What Is GLM-Legal?
1.1 The Model
GLM-Legal is a domain-specialized legal agent built through a three-stage training pipeline on top of GLM-5.3-Flash:
| Property | Value |
|---|---|
| Base model | GLM-5.3-Flash |
| Total parameters | 320 billion |
| Active parameters per token | 18 billion (MoE) |
| Architecture | Hybrid sparse + linear attention |
| Context window | 1,000,000 tokens |
| Max output | 128,000 tokens |
| Input modalities | Text, Image, Video, File |
| Output modalities | Text |
| Reasoning | Always enabled (low/high/max effort) |
| Pricing | $0.075/M input, $0.25/M output |
| Access | API beta (application required) |
| Platform | AutoClaw (coming soon) |
| Release date | October 12, 2026 |
The model inherits GLM-5.3-Flash's hybrid sparse and linear attention architecture, which reduces attention computation by 3.01× and KV cache by 4.44× compared to GLM-5.3, while maintaining long-context quality. This efficiency is critical for legal work, where agents often need to process hundreds of documents (contracts, filings, evidence) within a single context window.
1.2 The Z.ai Context
Z.ai (formerly Zhipu AI) has been building the GLM model family since 2021, evolving from research prototypes to commercial frontier models. The GLM-5.3 family, released in late August 2026, marked a significant step with native multimodal capabilities and aggressive pricing. GLM-Legal extends this trajectory into domain specialization:
2. Training Methodology: Three Stages to Legal Competence
Z.ai's approach to building GLM-Legal is noteworthy for its emphasis on training efficiency over parameter scale. The three-stage pipeline is:
2.1 Mid-Training: Domain Foundation
The mid-training phase expands the model's legal knowledge base by training on high-quality legal corpora that cover:
- Legal concepts and terminology
- Statutory rules and regulatory frameworks
- Interpretive frameworks and case law reasoning
- Deductive connections between facts, rules, and conclusions
This stage is designed to create the "knowledge substrate" required for complex downstream execution, rather than simply memorizing legal text.
2.2 Post-Training: Agent-Harness Co-Evolution
The post-training phase is where GLM-Legal differentiates itself from simpler fine-tuning approaches. Z.ai used reinforcement learning across tens of thousands of realistic legal environments where the model learns to:
- Plan multi-step legal analysis
- Advance workflows under concrete constraints
- Generate client-ready deliverables
Critically, the Harness (Z.ai's runtime orchestration framework) and the model are co-optimized:
"Edge cases surfaced during model training pinpoint architectural bottlenecks in the harness. Conversely, harness refinements expose the model to realistic execution constraints, training it to retrieve relevant context, adapt via intermediate execution signals, and meet production-grade standards from prompt ingestion to final sign-off."
This bi-directional learning ensures the model isn't just smart in isolation — it's effective within the full execution pipeline.
2.3 Production Feedback Loops
The third stage involves real-world practitioner feedback that drives continuous improvement. Legal professionals evaluate outputs against concrete criteria:
- Are all critical issues and edge cases addressed?
- Are arguments well-substantiated and citations legally sound?
- Does the deliverable strictly adhere to the matter brief?
- Can downstream teams use the output immediately without manual rework?
These findings are fed back into task environments, evaluation rubrics, and harness design for subsequent training cycles.
3. Benchmark Performance
3.1 Harvey Legal Agent Benchmark (LAB)
The Harvey LAB is the gold standard for evaluating legal agents, covering 24 practice areas with real-world law firm workflows under strict all-pass grading. GLM-Legal's performance:
| Model | Harvey LAB All-Pass | Rank (Vals) |
|---|---|---|
| Muse Spark 1.2 | 25.42% | #1 |
| Muse Spark 1.3 Max | 23.75% | #2 |
| Muse Spark 1.3 | 22.92% | #3 |
| Gemini 4 Argon | 19.58% | #5 |
| GLM-Legal | 19.17% | #7 |
| Kimi K3 | 12.9% | — |
| Claude Fable 5 | 11.3% | — |
| GLM-5.3-Flash (base) | 6.7% | — |
Key observations:
- GLM-Legal improved on its base model by 12.5 percentage points — a massive relative gain
- It ranks within 0.4 percentage points of Gemini 4 Argon, despite running on a significantly leaner parameter budget
- The results were independently evaluated by Vals (see vals.ai/models/zai_glm-legal)
3.2 LegalBench (Domain Reasoning)
LegalBench tests specialized legal reasoning across contracts, merger filings, and judicial opinions:
| Model | LegalBench Score |
|---|---|
| Claude Fable 5 | 88.6 |
| GPT-6 Astra | 87.1 |
| GLM-Legal | 86.2 |
| Kimi K3 | 86.2 |
| GLM-5.3-Flash (base) | 83.9 |
GLM-Legal achieves parity with Kimi K3 and trails the top models by only 1.9–2.4 points, despite the massive difference in model scale and training investment.
3.3 PLawBench (Chinese Legal Practice)
PLawBench evaluates Chinese legal consultation, case analysis, and document drafting under fine-grained rubrics:
| Model | PLawBench Score |
|---|---|
| GLM-Legal | 77.3 |
Z.ai claims this achieves "parity with leading frontier systems" on Chinese legal tasks. Given Z.ai's Chinese origins and the depth of Chinese legal corpora used in mid-training, this is a natural strength area.
3.4 Cost-Adjusted Performance
Perhaps the most striking metric is performance per dollar:
| Model | Harvey LAB | Cost (input/output per M) | Cost per 1% LAB point |
|---|---|---|---|
| GLM-Legal | 19.2% | $0.075 / $0.25 | ~$0.004 / ~$0.013 |
| Gemini 4 Argon | 19.6% | $2.00 / $10.00 | ~$0.102 / ~$0.510 |
| Meta Muse Wonder | ~13-15%* | $4.00 / $20.00 | ~$0.267 / ~$1.333 |
| Kimi K3 | 12.9% | $3.00 / $15.00 | ~$0.233 / ~$1.163 |
*Estimated based on Harvey LAB leaderboard positioning
GLM-Legal delivers ~25× better cost efficiency than Gemini 4 Argon on the Harvey LAB benchmark, making it particularly attractive for high-volume legal workflows.
4. Real-World Legal Workflows
Z.ai demonstrated GLM-Legal's capabilities across three concrete legal scenarios, each showing the gap between the base GLM-5.3-Flash and the specialized GLM-Legal:
4.1 Contract Review: Exit Clause Optimization
Task: Review a software services contract for a client switching providers, focusing on termination and handover.
GLM-5.3-Flash provided generic advice: "Consider shortening the notice period" and "require return of client data."
GLM-Legal delivered specific, actionable terms:
- Reduce convenience-termination notice from 60 to 30 days
- Remove cure periods for IP, confidentiality, or data-security breaches (immediate termination)
- Require return of all deliverables and work in progress, whether or not paid for
- Add up to 90 days of transition assistance covering source code, documentation, data, and knowledge transfer
The difference is the difference between general guidance and negotiation-ready provisions.
4.2 Evidence Analysis: Contradiction Mapping
Task: Analyze 40 documents from a software project dispute to identify contradictions in the provider's progress reports.
GLM-Legal connected external statements with conflicting internal records, cited specific source documents (BR-000024, BR-000027, BR-000014, BR-000015), referenced contract sections (MSA §§2 and 5), and explained the potential legal significance of each contradiction — while keeping facts and inferences distinct.
4.3 Compliance Research: Cross-Border Launch
Task: Plan privacy compliance for a website launching in Singapore, Malaysia, and Indonesia.
GLM-Legal broke requirements into prioritized action items with specific legal references:
- SG PDPA s26 (data transfer safeguards)
- MY PDPA s129 (overseas transfer requirements)
- ID PDP Law Arts 55–56 (cross-border data flow rules)
- Opt-in requirements for non-essential analytics
- Suppression list management for unsubscribe requests
5. Architecture: Why Lean Works for Legal
5.1 The Efficiency Advantage
GLM-Legal's underlying architecture — 320B total parameters with only 18B active — is a strategic advantage for legal work:
Legal agents often need to process hundreds of documents — contracts, filings, evidence, precedents — within a single session. The reduced KV cache means GLM-Legal can handle larger contexts at lower cost, making it practical for real-world legal workflows where context size is often the bottleneck.
5.2 The Specialization Multiplier
The key insight from GLM-Legal is that domain specialization can compensate for smaller scale. By focusing the model's training on legal tasks specifically:
- Mid-training builds deep legal knowledge without needing a larger base model
- Post-training with RL teaches the model to execute legal workflows, not just answer legal questions
- Harness co-optimization ensures the model works effectively within the full tool ecosystem
This approach contrasts with the "bigger is better" strategy of models like Gemini 4 Argon or Meta Muse Wonder Full, which rely on massive scale and general capability to handle legal work as one of many domains.
6. Comparison with Recent Legal AI Releases
GLM-Legal's release on October 12, 2026 sits in a crowded week of legal AI announcements:
| Model | Release Date | Harvey LAB | LegalBench | Base Model Scale | Pricing (in/out) |
|---|---|---|---|---|---|
| GLM-Legal | Oct 12, 2026 | 19.2% | 86.2 | 320B/18B (MoE) | $0.075/$0.25 |
| Gemini 4 Argon | Oct 13, 2026 | 19.6% | — | Not disclosed | $2.00/$10.00 |
| Meta Muse Wonder Full | Oct 12, 2026 | ~13.1% | — | Not disclosed | Subscription |
| Kimi K3 | Earlier 2026 | 12.9% | 86.2 | Not disclosed | $3.00/$15.00 |
| Claude Fable 5 | Earlier 2026 | 11.3% | 88.6 | Not disclosed | Varies |
See Gemini 4 Argon Global Developer Enterprise Rollout 1m Output Tokens Frontier Coding Cybersecurity Robotics 2026 10 14 for Gemini 4 Argon's full analysis and Meta Muse Wonder Wonder Full Frontier Agentic Coding Long Running Legal Active Memory 2026 10 13 for Meta's legal capabilities.
Notable patterns:
- GLM-Legal and Gemini 4 Argon are nearly tied on Harvey LAB (19.2% vs 19.6%), despite GLM-Legal's much smaller parameter count and dramatically lower cost
- LegalBench shows a different story: Claude Fable 5 leads at 88.6, suggesting that domain reasoning (single-turn) and agentic execution (multi-step) reward different capabilities
- The cost gap is enormous: GLM-Legal is 25–40× cheaper than its closest competitors on equivalent tasks
7. Deployment and Access
7.1 Current Access
GLM-Legal is currently in API beta with application-required access:
- Application form: z.ai/consultation/legal
- Target users: Law firms, legal tech companies, in-house legal teams, independent practitioners
- AutoClaw integration: Coming soon (Z.ai's desktop-native agent platform)
7.2 Integration Pattern
Based on the AutoClaw demos, GLM-Legal is designed to work within a full agent ecosystem:
The model supports function calling, streaming output, structured output (JSON), and context caching — all essential for integration into legal workflows.
7.3 Pricing Strategy
At $0.075/$0.25 per million tokens, GLM-Legal is positioned as a high-volume, cost-effective legal AI:
- A typical contract review (50K input tokens, 10K output tokens) costs approximately $3.75
- A full case analysis (500K input, 50K output) costs approximately $38.75
- Compare with Gemini 4 Argon: the same contract review would cost $110, and the case analysis $1,050
This pricing makes GLM-Legal viable for routine legal tasks at scale — contract review, compliance checks, document drafting — where cost per task matters more than marginal performance differences.
8. Key Takeaways
-
Specialization beats scale for domain tasks: GLM-Legal demonstrates that a lean 18B-active-parameter model, properly trained on domain-specific data and workflows, can compete with much larger frontier models on specialized tasks.
-
The legal AI race is intensifying: Within a single week (October 12–13), three major legal AI releases emerged (GLM-Legal, Gemini 4 Argon, Meta Muse Wonder Full), each with different strengths and pricing strategies.
-
Cost is becoming a differentiator: GLM-Legal's 25× cost advantage over comparable models makes it attractive for high-volume legal workflows, potentially democratizing access to frontier-tier legal AI.
-
Agent-harness co-optimization matters: Z.ai's approach of jointly optimizing the model and its runtime environment produces better results than model-only training, suggesting that the "agent stack" is as important as the model itself.
-
Chinese legal AI is maturing: GLM-Legal's strong PLawBench performance (77.3) and its Chinese-language training data represent a significant step for Chinese legal AI, complementing its strong English-language performance.
9. Limitations and Considerations
- Beta access only: GLM-Legal requires application for API access, limiting immediate adoption
- No public weights: Unlike some GLM models, GLM-Legal appears to be API-only (no open-weight release announced)
- Benchmark caveats: Harvey LAB has faced some criticism regarding task design and rubrics; treat leaderboard numbers as signals, not absolute truths
- Domain-specific: GLM-Legal excels at legal tasks but may not generalize well to other domains compared to general-purpose models
- Independent verification: While Vals independently evaluated Harvey LAB and LegalBench results, the PLawBench results are self-reported by Z.ai
10. Future Directions
Several trajectories emerge from GLM-Legal's release:
10.1 Open-Weight Legal Models
If Z.ai follows its pattern with other GLM models, an open-weight release of GLM-Legal (or a variant) could be forthcoming, enabling self-deployment for law firms with data sovereignty requirements.
10.2 Multi-Jurisdiction Expansion
The model's strong cross-border compliance capabilities (demonstrated in the Singapore/Malaysia/Indonesia example) suggest potential for expansion into more jurisdictions, particularly across Asia-Pacific where Z.ai has natural advantages.
10.3 The Specialization Wave
GLM-Legal may be the first of many domain-specialized models from Z.ai. Following the pattern of mid-training + post-training + harness co-optimization, we could expect similar models for finance, healthcare, or engineering.
10.4 Price Pressure on Frontier Models
At $0.075/$0.25 per million tokens, GLM-Legal sets a new price floor for legal AI. If it proves reliable in production, it could force competitors to lower prices or differentiate on capabilities beyond benchmarks.
11. References & Resources
Official Sources
- GLM-Legal: Advancing Agentic Legal Intelligence — Z.ai official announcement
- GLM-Legal on AutoClaw — Detailed capabilities and workflow demos
- GLM-5.3-Flash Documentation — Base model specifications
- Z.ai Pricing — Current API pricing
- GLM-Legal Beta Application — Apply for API access
Independent Evaluation
- Vals AI: GLM-Legal — Independent benchmark results
- Vals AI: Harvey LAB Leaderboard — Full legal agent rankings
Related Research
- Gemini 4 Argon Global Developer Enterprise Rollout 1m Output Tokens Frontier Coding Cybersecurity Robotics 2026 10 14 — Gemini 4 Argon's legal capabilities (19.6% Harvey LAB)
- Meta Muse Wonder Wonder Full Frontier Agentic Coding Long Running Legal Active Memory 2026 10 13 — Meta Muse Wonder Full's legal performance (13.1% Harvey LAB)
- Mistral Large 4 Le Chonk 1 Trillion Parameter Open Weight Frontier Moe Cybersecurity Agentic Coding 2026 10 12 — Mistral Large 4's MoE architecture comparison
GLM-Legal proves that in the legal AI race, a focused sprint on a lean foundation can keep pace with the heavyweights — and do it at a price that makes the difference between a pilot project and production deployment.