A thinking (reasoning) model in SB Intuitions' third Sarashina generation, added on 29 October 2026 to the free-beta Sarashina service that opened on 13 October. The blog states it was "developed fully from scratch" (フルスクラッチで開発), like Sarashina3 mini and nano, and that it reasons before answering, where Sarashina3 mini is an instruct-only model. Thinking depth is set per request with a reasoning_effort parameter (low / medium / high).

Pretraining uses the same Japanese, English and code-centred data as Sarashina3 mini, with long-context and agent-oriented data added for this version. Post-training is SFT, multi-task RL and OPSD (on-policy self-distillation) for style alignment, plus two new RL stages for search and for safety. Parameter count, architecture, token count, context window and license are not disclosed. The model is API-only, with no weights released.

The API documentation reports these self-reported scores at medium effort, against the non-thinking Sarashina3 mini: AIME 2025 0.789 (0.633), GPQA-Diamond 0.645 (0.571; 0.670 at high), SWE-bench Verified 0.326 (0.018), IFBench 0.695 (0.585), τ²-bench 0.554 (0.343), BFCL v3 0.749 (0.598), and SB Intuitions' Japanese agent benchmark J-tau 0.768 (0.385). Japanese cultural QA (JamC-QA) fell from 0.821 to 0.694. In the blog's normalised comparison against llm-jp-4.1-33b-thinking, Qwen3-32B, Qwen3-235B-A22B-Thinking-2507, gpt-oss-120b and GPT-5.4 mini, it scores highest on J-tau, BFCL v3 and the JMT-Safety multi-turn safety set. It trails Qwen3-235B-A22B-Thinking-2507 and GPT-5.4 mini on GPQA-Diamond and J-MultiChallenge. The authors note it was not specifically trained for coding-agent use. Artificial Analysis does not list it.

Model Details

Benchmark Scores

Benchmark Score Mode
AIME 2025 78.9% reasoning_effort medium
GPQA-Diamond 67.0% reasoning_effort high
SWE-bench Verified 34.0% reasoning_effort high
IFBench 72.9% reasoning_effort high
BFCL v3 75.0% reasoning_effort high
tau2-bench 55.4% reasoning_effort medium
japaneseproprietarysovereignreasoning

Related