Sarashina2.2

model
SB Intuitions

Compute-efficient Sarashina iteration focused on small, high-quality Japanese models. Dense Llama-style Transformers at 0.5B / 1B / 3B with both base and instruction-tuned variants. MIT licensed. Sarashina2.2-3B-Instruct scores 3.75 on Elyza-tasks-100 (vs 2.99 for Qwen2.5-3B-Instruct), 6.51 on Japanese MT-Bench, and 7.71 on English MT-Bench, demonstrating that a small Japanese-specialized model can outperform larger Japanese-capable models on Japanese tasks.

The 2.2 generation was extended into multimodal variants: Sarashina2.2-Vision-3B (image→text, November 2025) and Sarashina2.2-OCR (4B, ~March 2026, Japanese-document OCR).

HuggingFace (3B-Instruct)HuggingFace (Vision)HuggingFace (OCR)

Outputs 3

Sarashina2.2 (0.5B / 1B / 3B)

model

Small Japanese LLM family with base + instruct variants at 0.5B, 1B, 3B. MIT.

HuggingFace (3B-Instruct)
Architecture DENSE
Parameters 3B

Variants

Name Parameters Notes
Sarashina2.2-0.5B 0.5B —
Sarashina2.2-1B 1B —
Sarashina2.2-3B 3B —

Sarashina2.2-Vision-3B

model

Vision-language variant of Sarashina2.2-3B. MIT.

HuggingFace
Architecture DENSE
Parameters 3B

Sarashina2.2-OCR

model

4B OCR-specialized variant of the Sarashina2.2 family for Japanese documents.

HuggingFace
Architecture DENSE
Parameters 4B
open-weightjapanesemultimodal

Related