The Pulse
SB Intuitions Adds Reasoning and Tool Use to Sarashina
SB Intuitions introduced Sarashina3.2 mini thinking on October 29, adding adjustable reasoning, web search and multi-step tool use to its Japanese-developed model series. The company reports strong agent-benchmark results but also acknowled
AI.info Team · Written with AI · fact-checked ·
SB Intuitions says its new Sarashina model can handle harder tasks through reasoning and tool use, but the company’s own benchmarks show uneven results: the model leads on two agent tests while falling behind larger rivals on some science and multi-turn dialogue evaluations. The company introduced Sarashina3.2 mini thinking on October 29, describing it as a reasoning-focused addition to its Japanese-developed model series.
Reasoning depth can be set per request
Unlike Sarashina3 mini, which the company describes as an instruction model, Sarashina3.2 mini thinking generates an answer after analyzing the prompt and considering possible responses. Developers can choose low, medium or high reasoning effort through an API parameter called reasoning_effort, adjusting the balance between response time and the depth of processing.
SB Intuitions says it expanded long-context and agent-oriented training data to improve complex reasoning and agent performance. Its post-training process combines supervised learning, reinforcement learning across multiple tasks and on-policy self-distillation; the company also added reinforcement learning aimed at improving search and safety.
Web search and customer-support tools are central demos
The company illustrates web search with a question about an attack technique it says was absent from the model’s training data. Sarashina3.2 mini thinking searches the web and answers with sources, while refusing to provide instructions for reproducing the attack. The example shows the company’s intended distinction between using external information to explain a subject and withholding potentially harmful operational guidance.
A second demonstration uses an online retailer’s customer-support system. Across multiple turns, the model checks an order, processes a cancellation request and applies the service’s cancellation policy using information returned by tools. SB Intuitions says the example worked with reasoning effort set to low because the order status and relevant policy came directly from the system.
The post also shows the model generating HTML and JavaScript to render a torus knot and a Clifford attractor, and extracting unstructured information into JSON or a Markdown table. Those examples support the company’s case for using the model in software and workflow tasks, though they are demonstrations prepared by SB Intuitions, not independent tests.
Agent scores rise, but gaps remain
SB Intuitions reports evaluations across 12 benchmarks covering dialogue, Japanese cultural knowledge, mathematics, coding, STEM, instruction following, agent tasks, search and safety. The comparison set includes the company’s Sarashina3 mini, llm-jp-4.1-33b-thinking from Japan’s National Institute of Informatics, Alibaba’s Qwen3 models and OpenAI’s gpt-oss-120b and GPT-5.4 mini.
The company says Sarashina3.2 mini thinking scored highest among the compared models on J-tau and BFCL v3, two tests of agent and function-calling ability, and performed at roughly the same level as the top comparison on tau2-bench. It also reports gains over Sarashina3 mini in multi-turn dialogue, mathematics, agent tasks, search and safety. The authors summarize their results this way:
以上のように、Sarashina3.2 mini thinking は、エージェント能力を中心に幅広いドメインで性能向上を達成しました。
浅野広樹、泉健太、金輝燦、澁谷紘人, SB Intuitions
The company also identifies limits. Sarashina3.2 mini thinking scored lower than Sarashina3 mini on JamC-QA, a Japanese culture question-and-answer test, and trailed Qwen3-235B-A22B-Thinking-2507 and GPT-5.4 mini on GPQA-Diamond, a STEM benchmark, and J-MultiChallenge, a multi-turn dialogue test. SB Intuitions adds that it has not done enough training specifically for using the model as a coding agent.
The company says it used official implementations for J-tau, tau2-bench and BFCL v3, while it ran the other evaluations with internally implemented evaluation code using FlexEval. It normalized benchmark scores to a 0-to-1 range and set reasoning effort to medium where a model allowed that setting. The results are company-reported comparisons, not an independent assessment.
The model joins Sarashina’s free beta service
SB Intuitions says Sarashina3.2 mini thinking is being added to the Sarashina service, which offers API access during a free beta. The service launched on October 13 and is available to developers, IT engineers and individual users; the company says usage is subject to rate limits. It has not set an end date for the beta.
The release puts adjustable reasoning and system actions alongside the series’ earlier text-generation and other model APIs. For the new model, the company’s published evaluations point to its strongest case in agent tasks—and spell out where its scores still fall short.