Swallow - Post-trained Japanese - WMT20 Ja-En: leaderboard
From the Swallow LLM Leaderboard (Institute of Science Tokyo) post-trained Japanese board, which evaluates instruction-tuned and reasoning models with the Swallow evaluation harness. Japanese-to-English news translation from WMT20, zero-shot with a chain-of-thought prompt, scored with corpus BLEU (0-100, higher is better).
Metric: BLEU. Source: swallow-llm.github.io. Saturation forecast: Around 2031. 68 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) | 26.5 | #343 |
| 2 | gemma-4-26B-A4B-it (Thinking) | 26.1 | |
| 3 | GPT-4.1 | 26 | #258 |
| 4 | Gemma 4 31B (IT) (Thinking) | 26 | #178 (Gemma 4 31B (IT)) |
| 5 | Llama-3.3-Swallow-70B-Instruct-v0.4 | 25.1 | #540 |
| 6 | GPT-5.4 (2026-03-05) (Medium) | 24.9 | #63 (GPT-5.4 (2026-03-05)) |
| 7 | Qwen 3.5 122B A10B (Thinking) | 24.7 | #184 (Qwen 3.5 122B A10B) |
| 8 | Llama 3.3 70B Instruct | 24.6 | #541 |
| 9 | Qwen 3.5 27B (Thinking) | 24.1 | #192 (Qwen 3.5 27B) |
| 10 | Qwen 3.5 35B A3B (Thinking) | 23.9 | #268 (Qwen 3.5 35B A3B) |
| 11 | GPT-5 Mini (2025-08-07) (Medium) | 23.9 | #175 (GPT-5 Mini (2025-08-07)) |
| 12 | Gemma 3 27B (IT) | 23.8 | #528 |
| 13 | GPT-5 (Medium) | 23.6 | #105 (GPT-5) |
| 14 | Llama 4 Scout Instruct | 23 | #547 |
| 15 | Qwen 3 235B A22B 2507 Instruct | 23 | #315 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swallow-post-trained-japanese-wmt20-ja-en · How It Works · Data refreshed daily, snapshot 2026-11-05.