Swallow - Japanese MT-Bench - Roleplay: leaderboard

Metric: Judge Score (normalized, %). Source: swallow-llm.github.io. 68 models tracked.

Top models

#ModelScore
1GPT-5 (Medium)82.8
2GPT-5.4 (Medium)81.6
3GPT-5 Mini (Medium)81.2
4O3 (2025-04-16) (Medium)79.6
5gemma-4-26B-A4B-it (Thinking)75.3
6Gemma 4 31B (IT) (Thinking)75
7GPT-4.174.8
8Qwen 3.5 122B A10B (Thinking)73.1
9O3 Mini (Medium)72.8
10Gemma 4 E4B (IT) (Thinking)71.9
11Qwen 3 235B A22B 2507 Instruct71.6
12GPT-OSS-Swallow-120B-RL-v0.1 (Thinking)71.6
13Gemma 3 27B (IT)70.5
14Qwen 3 Next 80B A3B (Thinking)70.5
15GPT-OSS-120B (Medium)70.4

Interactive version: theaggregate.ai/benchmark?slug=swallow-japanese-mt-bench-roleplay · How It Works · Data refreshed daily, snapshot 2026-09-19.