TRACE Bench - Language Quality: leaderboard
TRACE Bench per-turn language quality: an LLM judge checks each roleplay reply for fluency, grammar, usage and internal-logic problems, without scoring persona, length or diversity (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent over scorable replies; higher is better.
Metric: Per-turn LLM-judge language quality (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiMo-V2.5 | 99.03 | #196 |
| 2 | MiMo-V2.5-Pro | 98.73 | #167 |
| 3 | DeepSeek V4 Flash | 98.58 | #155 |
| 4 | Claude Sonnet 4.6 | 98.16 | #94 |
| 5 | Kimi K2.5 (Non-reasoning) | 98.16 | #152 (Kimi K2.5) |
| 6 | GPT-5.5 | 97.87 | #33 |
| 7 | Gemma 4 31B (Non-reasoning) | 97.66 | #212 (Gemma 4 31B) |
| 8 | DeepSeek V4 Pro (Non-reasoning) | 97.41 | #112 (DeepSeek V4 Pro) |
| 9 | Qwen 3 8B (Non-reasoning) | 97.27 | #678 (Qwen 3 8B) |
| 10 | Qwen 3 14B (Non-reasoning) | 96.98 | #542 (Qwen 3 14B) |
| 11 | Qwen 3.5 27B (Non-reasoning) | 96.59 | #192 (Qwen 3.5 27B) |
| 12 | MiniMax-M2.7 | 96.53 | #270 |
| 13 | GLM-5 (Non-reasoning) | 95.51 | #149 (GLM-5) |
| 14 | Mistral Small 3.2 | 90.77 | #608 |
| 15 | Hermes-4-14B | 84.91 | #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-language-quality · How It Works · Data refreshed daily, snapshot 2026-11-05.