SB Intuitions launches Sarashina3.2 mini thinking, the Sarashina series' first reasoning model, via API

SB Intuitions announced Sarashina3.2 mini thinking on October 29, the first "thinking model" in its Sarashina series. The model is available by API through the company's Sarashina service, which is currently running as a free beta. The announcement was reported in Japanese by ITmedia AI+ on October 30, 2026.

SB Intuitions is a Tokyo-based AI developer (Minato ward) under the SoftBank group. It develops the Sarashina series from scratch. A thinking model, as ITmedia describes it, spends more time generating output in order to raise answer accuracy.

What is new

• Reasoning tier: Sarashina3.2 mini thinking can handle more complex tasks than Sarashina3 mini, the series' non-thinking model.
• Capability gains: The company says coding, tool calling and web search performance improved over Sarashina3 mini.
• Adjustable depth: Users can set "thinking depth" to low, medium or high, trading response time against accuracy.
• Training focus: Pre-training targeted long-text processing and agent performance. Post-training focused on web search ability and safety.

Use cases and benchmarks

SB Intuitions points to two use cases: answering questions grounded in web search, and operating systems built around e-commerce customer support.

The company claims the model scores comparably on multiple benchmarks to three reference models:

• llm-jp-4.1-33b-thinking from Japan's National Institute of Informatics
• Qwen3-235B-A22B-Thinking-2507 from China's Alibaba Cloud
• gpt-oss-120b from OpenAI

These results are company claims. Specific benchmark names and scores are not given in this announcement coverage.

Availability

The model is served through the Sarashina service via API. Pricing not yet disclosed. Context window and parameter count are also not yet disclosed.

What this means

Sarashina3.2 mini thinking adds a reasoning option to a series that previously had only non-thinking models, and it targets the workloads SB Intuitions names: tool calling, web search and system operation. The three-level depth setting lets developers choose a latency and accuracy trade-off per request instead of committing to one behavior. The comparison set, which includes a Japanese research-institute model, an Alibaba Cloud model and an OpenAI model, shows where SB Intuitions positions the release. Without published scores, pricing and context length, developers cannot yet judge those claims independently, and they will need to test the model against their own tasks through the free beta.