GLM-4.7: benchmark results
Zhipu AI's GLM-4.7, a 355B sparse MoE (32B active) flagship reasoning model (December 2025). Provider: Zhipu. Released 2025-12-22. Access: Open.
Aggregate Rating 1702 ± 4, rank #192 of 1656 rated models, from 922 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BALSAM - Sequence Tagging | 46.49 | Overall score (0-100, LLM-judged generation and multiple cho | 100 |
| CryptoAnalystBench - Fabricated Claims | 1.22 | Fabricated claims (%): share of the factual claims extracted | 100 |
| DV-World - DV-Sheet Dashboard | 38.03 | Dashboard overall (%): MLLM rubric score (insightfulness, ac | 100 |
| Drift-Bench (Service-Oriented, No Clarification) - Intention | 69.89 | Success rate (%; 150 StableToolBench G1 API tasks with an in | 100 |
| Drift-Bench (Service-Oriented, No Clarification) - Premise | 56.65 | Success rate (%; 150 StableToolBench G1 API tasks with a pre | 100 |
| Drift-Bench (State-Oriented, No Clarification) - Expression | 50 | Success rate (%; 200 AgentBench operating-system and databas | 100 |
| LongJud-Bench - Evidence Marshalling (Second Instance) | 71 | Evidence Marshalling Score (%; Claude Sonnet 4.6 judge, 0-10 | 100 |
| OpenCompass Agent - Tool Use (CompassBench 2601) | 68.8 | Score (%) | 100 |
| RT-QA - Direct (L1) | 46.9 | Accuracy (%) on direct (0-hop) questions, mean of the six da | 100 |
| SWE-rebench V2 - Go | 17.2 | pass@1 (%, 0-100) on the 60 Go tasks, averaged over three ru | 100 |
| AGC-Bench - historical_analogy | 1.94 | Dataset z-score | 99.4 |
| AA LiveCodeBench | 89.42 | Pass@1 (%) | 99 |
Interactive version: theaggregate.ai/model?slug=glm-4-7 · How It Works · Data refreshed daily, snapshot 2026-10-20.