HardMTBench: leaderboard
General-purpose machine translation benchmarks such as FLORES-200 have reached a saturation regime on Chinese-English pairs, where modern large language models cluster within a narrow band of high scores. Across 22 systems, FLORES-200 zh-en GEMBA scores fall in a 7.87-point range with a standard deviation of 2.29, which compresses the separation between systems on knowledge-intensive domains such
Metric: HardMTBench zh-en GEMBA-DA (self-reported). Source: benchmarklist.com. Saturation forecast: Around June 2027. 22 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 | 91.4 | #31 |
| 2 | GPT-5.5 (Non-reasoning) | 91.27 | #31 (GPT-5.5) |
| 3 | DeepSeek V4 Pro | 90.91 | #102 |
| 4 | Qwen 3.6 35B A3B | 90.68 | #226 |
| 5 | Qwen 3.5 397B A17B (Non-reasoning) | 90.66 | #146 (Qwen 3.5 397B A17B) |
| 6 | Gemma 4 31B | 90.62 | #200 |
| 7 | DeepSeek V4 Pro (Non-reasoning) | 90.59 | #102 (DeepSeek V4 Pro) |
| 8 | Qwen 3.5 397B A17B | 90.47 | #146 |
| 9 | Gemini 3.1 Pro (Preview) | 90.24 | #58 |
| 10 | Gemma 4 26B A4B | 89.65 | #358 |
| 11 | Gemma 4 31B (Non-reasoning) | 89.3 | #200 (Gemma 4 31B) |
| 12 | Qwen 3.6 35B A3B (Non-reasoning) | 89.03 | #226 (Qwen 3.6 35B A3B) |
| 13 | Hy-MT2-30B-A3B | 88.6 | #359 |
| 14 | Gemma 4 26B A4B (Non-reasoning) | 87.82 | #358 (Gemma 4 26B A4B) |
| 15 | Gemma 4 E4B | 82.69 | #767 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=hardmtbench · How It Works · Data refreshed daily, snapshot 2026-10-20.