TRACE Bench - Length: leaderboard
TRACE Bench reply-length compliance: a deterministic check of whether each roleplay reply falls within the appropriate length range, using word counts for English and character counts for Chinese or mixed replies (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent; higher is better; an auxiliary dimension.
Metric: Reply-length compliance (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 14B (Non-reasoning) | 99.6 | #542 (Qwen 3 14B) |
| 2 | Gemma 4 31B (Non-reasoning) | 98.49 | #212 (Gemma 4 31B) |
| 3 | Qwen 3 8B (Non-reasoning) | 98.32 | #678 (Qwen 3 8B) |
| 4 | GPT-5.5 | 96.54 | #33 |
| 5 | MiniMax-M2.7 | 94.93 | #270 |
| 6 | MiMo-V2.5 | 94.59 | #196 |
| 7 | Kimi K2.5 (Non-reasoning) | 94.06 | #152 (Kimi K2.5) |
| 8 | Claude Sonnet 4.6 | 93.7 | #94 |
| 9 | Mistral Small 3.2 | 91.44 | #608 |
| 10 | DeepSeek V4 Pro (Non-reasoning) | 91.42 | #112 (DeepSeek V4 Pro) |
| 11 | MiMo-V2.5-Pro | 88.75 | #167 |
| 12 | DeepSeek V4 Flash | 88.7 | #155 |
| 13 | GLM-5 (Non-reasoning) | 86.38 | #149 (GLM-5) |
| 14 | Qwen 3.5 27B (Non-reasoning) | 81.41 | #192 (Qwen 3.5 27B) |
| 15 | Hermes-4-14B | 76.7 | #789 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trace-bench-length · How It Works · Data refreshed daily, snapshot 2026-11-05.