Swallow - Pre-trained Japanese - MGSM: leaderboard
From the Swallow LLM Leaderboard (Institute of Science Tokyo) pre-trained Japanese board, which evaluates base models without post-training. MGSM, the Japanese translation of GSM8K grade-school math word problems, 4-shot; metric is exact match. Upstream publishes 0-1 scores; shown as 0-100, higher is better.
Metric: Exact Match (%). Source: swallow-llm.github.io. Saturation forecast: Estimated already saturated. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B | 82.8 | #512 |
| 2 | Qwen 2.5 32B | 80.8 | #544 |
| 3 | Qwen3-14B-Base | 76.8 | #628 |
| 4 | Llama 4 Scout Base | 76 | #708 |
| 5 | Qwen 2.5 14B | 74 | #716 |
| 6 | Qwen3-30B-A3B-Base | 74 | #597 |
| 7 | Qwen3-8B-Base | 71.6 | #626 |
| 8 | gemma-3-27B-pt | 70.4 | #505 |
| 9 | Llama 3.1 70B | 67.2 | #588 |
| 10 | Qwen3-4B-Base | 64.4 | #880 |
| 11 | Qwen 2.5 7B | 61.6 | #895 |
| 12 | Gemma 2 27B | 59.6 | #803 |
| 13 | sarashina2.2-3B | 59.6 | #766 |
| 14 | gemma-3-12B-pt | 58.4 | #686 |
| 15 | plamo-2-8B | 50.8 | #839 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=swallow-pre-trained-japanese-mgsm · How It Works · Data refreshed daily, snapshot 2026-11-05.