Swallow - Japanese MT-Bench - Writing: leaderboard
From the Swallow LLM Leaderboard: Japanese MT-Bench (Nejumi edition, v4 questions with corrected v2 reference answers), two-turn dialogues graded by GPT-5.2 against references; normalized 0-1 upstream, shown as 0-100, higher is better. This board covers the Writing category: blog posts, emails and fiction.
Metric: Judge Score (normalized, %). Source: swallow-llm.github.io. Saturation forecast: Around February 2027. 68 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (2026-03-05) (Medium) | 80.4 | #63 (GPT-5.4 (2026-03-05)) |
| 2 | GPT-5 (Medium) | 78.1 | #105 (GPT-5) |
| 3 | GPT-5 Mini (2025-08-07) (Medium) | 77.7 | #175 (GPT-5 Mini (2025-08-07)) |
| 4 | Gemma 4 31B (IT) (Thinking) | 74.8 | #178 (Gemma 4 31B (IT)) |
| 5 | O3 (2025-04-16) (Medium) | 74.7 | #125 (O3 (2025-04-16)) |
| 6 | gemma-4-26B-A4B-it (Thinking) | 74.5 | |
| 7 | GPT-4.1 | 72.5 | #258 |
| 8 | Qwen 3 235B A22B 2507 Instruct | 71.7 | #315 |
| 9 | GPT-OSS-Swallow-120B-RL-v0.1 (Thinking) | 70.7 | #309 (GPT-OSS-Swallow-120B-RL-v0.1) |
| 10 | Qwen 3.5 27B (Thinking) | 70.2 | #192 (Qwen 3.5 27B) |
| 11 | Gemma 4 E4B (IT) (Thinking) | 69.7 | |
| 12 | O3 Mini (Medium) | 69.4 | #282 (O3 Mini) |
| 13 | Gemma 4 E2B (IT) (Thinking) | 69.2 | #779 (Gemma 4 E2B (IT)) |
| 14 | Qwen 3 235B A22B 2507 (Thinking) | 68.5 | #263 |
| 15 | GPT-OSS-120B (Medium) | 68 | #346 (GPT-OSS-120B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swallow-japanese-mt-bench-writing · How It Works · Data refreshed daily, snapshot 2026-11-05.