Sarashina2.1-1B

Sarashina2.1-1B is a compact 1-billion-parameter language model from SB Intuitions, SoftBank's in-house AI research company, released in 2024. It uses a Llama architecture with an 8,192-token context and was trained on a large web-scale Japanese and English corpus reported at around 10 trillion tokens. It is part of SoftBank's Sarashina program, the company's homegrown generative-AI effort built on NVIDIA NeMo and Megatron-LM tooling, which later produced much larger Sarashina models and the commercial Sarashina Mini API service.

The model is open for research under the Sarashina Model NonCommercial License Agreement, so commercial users must contact SB Intuitions for terms. It is best seen as a small, efficient Japanese-capable base for edge or cost-sensitive workloads rather than a frontier model.