Sarashina3
modelYour notes
SB Intuitions' third-generation flagship series, launched via SoftBank's sovereign cloud "Cloud PF Type A" (Oracle Alloy on domestic data centers): five models — mini (standard), nano, guard (safety classifier), embedding, and rerank. Pretrained from scratch on "30T tokens of large-scale, high-quality data" with curriculum learning; post-training uses multi-stage asynchronous RL plus OPSD (on-policy self-distillation). Evaluated against Qwen3 4B/30B/235B, GPT-4.1, and GPT-5.4 mini on eight benchmarks including ELYZA-tasks-100, AIME 2025, IFBench, BFCL v3, and τ²-bench Telecom. API-only; parameter counts undisclosed.
On 13 October 2026 SB Intuitions opened the Sarashina service (sarashina.ai) as a free public beta, with API and web-chat access to six Sarashina3 models: mini, nano, embedding, rerank, guard and a translate endpoint. A from-scratch reasoning successor, Sarashina3.2 mini thinking, was added on 29 October.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| Sarashina3 mini | — | — |
| Sarashina3 nano | — | — |
| Sarashina3 guard | — | — |
| Sarashina3 embedding | — | — |
| Sarashina3 rerank | — | — |
| Sarashina3 translate | — | Translation endpoint (/v1/translate); one of the six models in the Sarashina service free beta opened 2026-10-13 |