TRACE Bench - Length: leaderboard
Metric: Reply-length compliance (%). Source: arxiv.org. 26 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3 14B (Non-reasoning) | 99.6 |
| 2 | Gemma 4 31B (Non-reasoning) | 98.49 |
| 3 | Qwen 3 8B (Non-reasoning) | 98.32 |
| 4 | GPT-5.5 | 96.54 |
| 5 | MiniMax-M2.7 | 94.93 |
| 6 | MiMo-V2.5 | 94.59 |
| 7 | Kimi K2.5 (Non-reasoning) | 94.06 |
| 8 | Claude Sonnet 4.6 | 93.7 |
| 9 | Mistral Small 3.2 | 91.44 |
| 10 | DeepSeek V4 Pro (Non-reasoning) | 91.42 |
| 11 | MiMo-V2.5-Pro | 88.75 |
| 12 | DeepSeek V4 Flash | 88.7 |
| 13 | GLM-5 (Non-reasoning) | 86.38 |
| 14 | Qwen 3.5 27B (Non-reasoning) | 81.41 |
| 15 | Hermes-4-14B | 76.7 |
Interactive version: theaggregate.ai/benchmark?slug=trace-bench-length · How It Works · Data refreshed daily, snapshot 2026-09-19.