pfgen-bench - Chat Mode - Fluency: leaderboard
pfgen-bench (Preferred Networks) asks 50 Japanese questions expecting formal answers of about 100 characters and scores them without an LLM judge, averaged over trials; each metric is discounted linearly past 100 characters, to zero at 150. Chat mode: the instruction is a system prompt in a chat template. Fluency is 10-gram overlap with several reference answer sets, scaled so a typical reference scores 1; higher is better.
Metric: Fluency Score. Source: github.com. Saturation forecast: Estimated already saturated. 35 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 4 (20250514) (Thinking) | 0.97 | #164 (Claude Opus 4 (20250514)) |
| 2 | MiniMax-Text-01 | 0.86 | #519 |
| 3 | Qwen 2.5 72B Instruct | 0.85 | #469 |
| 4 | Mistral Large 2 (Nov) Instruct (2411) | 0.83 | #480 |
| 5 | QwQ 32B-Preview | 0.73 | #671 |
| 6 | sarashina2.2-3B-instruct-v0.1 | 0.72 | #1115 |
| 7 | Phi-4 | 0.71 | #707 |
| 8 | Mistral Small 3 | 0.64 | #630 |
| 9 | Qwen 2.5 14B Instruct | 0.64 | #654 |
| 10 | Qwen 3 8B (Non-reasoning) | 0.64 | #678 (Qwen 3 8B) |
| 11 | Qwen 2.5 32B Instruct | 0.63 | #509 |
| 12 | Qwen 2.5 7B Instruct | 0.54 | #863 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=pfgen-bench-chat-mode-fluency · How It Works · Data refreshed daily, snapshot 2026-11-05.