OpenModelsHQ

Japan

Open models from Japan

We track 15 open-weight models from 1 lab in Japan. They account for 1 of 205 tracked releases (0%) from Dec 2025 to Nov 2026. The newest is RePo-OLMo2-1B-stage2-L5, listed on 14 Dec 2025.

Explore the models
Minato, Tokyo
Minato, Tokyo Photo credit

Latest models

Newest listed first
  1. RePo-OLMo2-1B-stage2-L5Sakana AI · 1.5B
  2. ctm-maze-largeSakana AI · 32M
  3. ctm-imagenetSakana AI · 186M
  4. RLT-7BSakana AI · 7.6B
  5. Llama-3-Karamaru-v1Sakana AI · 8B
  6. TinySwallow-1.5B-InstructSakana AI · 1.5B
More models from Japan
  1. TinySwallow-1.5BSakana AI · 1.5B
  2. TAID-VLM-2BSakana AI · 2.2B
  3. TAID-LLM-1.5BSakana AI · 1.5B
  4. Llama-3-8B-Instruct-Coding-ExpertSakana AI · 8B
  5. Evo-Nishikie-v1Sakana AI · 1.3B
  6. DiscoPOP-zephyr-7b-gemmaSakana AI · 8.5B

Cheapest tracked APIs

No hosted API prices recorded. The lab pages link to the available weights.

How the work developed

Sakana AI published evolutionary model merging as a method for combining the capabilities of existing models. Its Japanese-language experiments combined language and mathematical capabilities, and also explored Japanese vision-language models. The team released models and research material describing how it selected the combinations.

TinySwallow extended Sakana’s work on Japanese language models through knowledge distillation. The model card names Qwen models as teacher and student, followed by instruction tuning for Japanese conversation. It also lists the training datasets and notes that users must observe both the source-model licence and the terms attached to Gemma-derived data. This is a record of the Japanese lab’s work, including its use of models and data developed elsewhere.

Sources and counting method

Counts cover selected repositories, including model variants. Dates come from Hugging Face repository creation dates and may differ from public announcements. Location follows the lab record, not the nationality of individual contributors.

Sources