Llama-3.3-Swallow-70B-Instruct-v0.4: benchmark results
Tokyo Tech's Llama 3.3 Swallow 70B Instruct v0.4, continually pretrained from Llama 3.3 70B on about 315B tokens of Japanese web text and instruction-tuned via SFT for Japanese. Provider: Institute of Science Tokyo. Released 2025-03-10. Access: Open.
Aggregate Rating 1574 ± 17, rank #529 of 1656 rated models, from 195 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Japanese LLM - Jcommonsenseqa Exact Match | 97.86 | Score (%) | 100 |
| Open Japanese LLM - AIO Char F1 | 89.25 | Score (%) | 99.5 |
| Open Japanese LLM - ALT E TO J Comet Wmt22 | 91.61 | Score (%) | 98.3 |
| Open Japanese LLM - Niilc Char F1 | 67.39 | Score (%) | 97.8 |
| Open Japanese LLM - MT | 85.99 | Score (%) | 97.6 |
| Open Japanese LLM - ALT J TO E Comet Wmt22 | 89.23 | Score (%) | 97.5 |
| Open Japanese LLM - QA | 72.74 | Score (%) | 97.1 |
| Open Japanese LLM - ALT E TO J Bleu JA | 14.94 | Score (%) | 97 |
| Open Japanese LLM - Wikicorpus J TO E Comet Wmt22 | 77.15 | Score (%) | 96.9 |
| Open Japanese LLM - Wikicorpus E TO J Comet Wmt22 | 85.96 | Score (%) | 96.6 |
| Open Japanese LLM - ALT E TO J Bert Score JA F1 | 87.44 | Score (%) | 96.2 |
| Open Japanese LLM - Wiki PAS SET F1 | 11.64 | Score (%) | 96 |
Interactive version: theaggregate.ai/model?slug=llama-3-3-swallow-70b-instruct-v0-4 · How It Works · Data refreshed daily, snapshot 2026-10-20.