Persian Fluency Bench - Context Retention: leaderboard
Whether Persian Fluency Bench responses hold the details established earlier in the conversation. Higher is better. Source: github.com/mshojaei77/ParsiEval.
Metric: Judge Score (%, judge google/gemini-2.5-flash). Source: github.com. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemma 3 12B (IT) | 85.9 | #672 |
| 2 | Qwen 3 30B A3B 2507 Instruct | 78.5 | #492 |
| 3 | Gemini 2.5 Flash | 76.7 | #255 |
| 4 | Qwen 3 Next 80B A3B Instruct | 76.6 | #381 |
| 5 | Gemma 3 4B (IT) | 76.5 | #997 |
| 6 | Gemma 3 27B (IT) | 73.5 | #528 |
| 7 | Gemma 3n E4B (IT) | 73.1 | #747 |
| 8 | granite-4.0-h-micro | 32.3 | #1078 |
| 9 | Qwen 3 4B | 17.1 | #834 |
| 10 | Llama 3.2 3B Instruct | 16.6 | #1327 |
| 11 | Mixtral 8x7B Instruct | 15 | #1085 |
| 12 | Ministral 3B | 13.9 | #1273 |
| 13 | lfm-2.2-6B | 4.1 | |
| 14 | LFM2 8B A1B | 3.9 | #1248 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=persian-fluency-bench-context-retention · How It Works · Data refreshed daily, snapshot 2026-11-05.