모델 목록으로
Preferred Networks조건부 오픈

PLaMo 2.0

The latest version of the domestic large language model developed by Preferred Networks (PFN). With 31 billion parameters, it achieves performance comparable to GPT-4 mini and Claude 2.5 in Japanese language tasks. Through a joint project with Sakura Internet and NICT, it leverages over 70 billion pages of Japanese web data accumulated by NICT for training.

파라미터

31B

컨텍스트

32K

라이선스

PLaMo License

출시일

2026-03-01

일본어 처리 능력

🇯🇵Native JP

Model developed by a Japanese company or specialized for Japanese. Highest Japanese understanding and generation capability.

API 가격

입력 가격 (1M 토큰당)

¥50

출력 가격 (1M 토큰당)

¥200

과금 모드: standard

강점

  • Extremely high Japanese language processing ability
  • Domestic model optimized for Japanese context
  • Trained on over 700 billion Japanese pages
  • API available via Sakura Internet

약점

  • Lags behind frontier models in global benchmarks
  • Shorter context length of 32K
  • Non-commercial license (commercial use requires consultation)
  • Slightly slow inference speed

활용 사례

  • Japanese document generation and summarization
  • Customer support (in Japanese)
  • AI systems compliant with Japanese regulations
  • AI operation in domestic cloud environments

심층 분석

Jaster Benchmark (4-shot, acc)

0.665

Highest among 31B-class models, surpassing gpt-4o-mini (0.635)

pfgen-bench (Japanese fluency)

0.890

Top score for 31B models, superior to gpt-4o-mini (0.804)

M-IFEval Japanese (avg)

0.677

Best instruction-following in Japanese among 31B models

Input Price

¥60/M tokens

~$0.40/M tokens, 75% cheaper than PLaMo 1.0 Prime

Context Length

32,000 tokens

2x increase from previous version for long-document processing

Token Efficiency (Japanese)

+45% improvement

Custom tokenizer reduces cost per character compared to standard tokenizers

강점

  • Outstanding Japanese language fluency and cultural nuance understanding, achieving top scores on Japanese-specific benchmarks.
  • Exceptional training efficiency through pruning, where an 8B model rivals the performance of its 100B predecessor.
  • Cost-effective deployment with optimized tokenizer and quantization support, significantly reducing API pricing.

약점

  • Lags behind in complex multi-step mathematical reasoning compared to specialized reasoning models.
  • Code generation performance, while improved, still trails behind models like Qwen3-8B-Base in coding benchmarks.
  • Long-context retrieval capabilities required architectural modifications (transitioning to full attention), indicating initial limitations of the hybrid Samba design.

경쟁사 비교

ModelArenaSWEGPQAPrice
PLaMo 2.0-31BN/AN/AN/A¥60/¥250 per M tokens
Qwen2.5-32B-InstructN/AN/AN/AOpen-source
gpt-4o-miniN/AN/AN/A$0.15/$0.60 per M tokens

PLaMo 2.0 represents a significant advancement in Japanese-focused large language models from Preferred Networks. The series features a hybrid architecture that initially combines Mamba's state space model with sliding window attention (Samba) for computational efficiency, then transitions to full attention via continual pre-training to overcome long-context retrieval limitations. This approach, coupled with innovative training techniques like extensive synthetic data generation and efficient model pruning, allows the 31B-parameter model to deliver performance comparable to its 100B-parameter predecessor while being far more resource-efficient.

The model demonstrates state-of-the-art results across Japanese benchmarks, excelling in language fluency (pfgen-bench), instruction-following (M-IFEval Japanese), and knowledge assessment (Jaster). Its commercial version, PLaMo 2.0 Prime, introduces a custom tokenizer that improves Japanese token efficiency by 45%, doubles the context length to 32,000 tokens, and reduces API pricing by over 75% compared to its predecessor. While the model shows some limitations in complex mathematical reasoning and code generation compared to specialized competitors, it stands as a premier choice for Japanese language applications, with successful deployment in services like QommonsAI and Tachyon AI.

분석 생성일: 2026-05-23