SoftBank's SB Intuitions Adds Sarashina3.2 mini thinking, a Japanese Reasoning LLM With Adjustable Thinking Depth
Primary sources: SB Intuitions TECH BLOG, "国産LLM Sarashina3.2 mini thinking の紹介" (Introducing the domestic LLM Sarashina3.2 mini thinking; published Oct 29, 2026, 13:30 JST, which is Oct 29, 12:30 AM ET), and the official Sarashina API docs and model performance page (checked Oct 30, 2026). Quotes are our translations from Japanese.
Oct 30, 2026: Japan's push for homegrown AI just got a model that thinks before it answers. SB Intuitions, the wholly owned SoftBank Corp. subsidiary (company profile: SoftBank Corp. 100% shareholder) building the Japanese-focused Sarashina LLM family, has added Sarashina3.2 mini thinking to its Sarashina service. Per the Sarashina API docs, it's currently the only "thinking" (reasoning) model on the API, and you can control how hard it thinks.
What's new
According to SB Intuitions' Oct 29 blog post:
- It's a thinking model. Sarashina3 mini is an instruct model that answers directly. Sarashina3.2 mini thinking instead goes through a reasoning process first (analyzing the input, listing candidate answers and checking them) before giving its final answer.
- Three reasoning levels. You choose
low,mediumorhighwith thereasoning_effortAPI parameter, trading speed against accuracy. - Built from scratch. Like Sarashina3 mini and nano, the company says it was developed "full scratch" (not fine-tuned from someone else's base model), pre-trained mainly on Japanese, English and code. This time the training data was boosted with long-context and agent-focused data.
- New RL stages. Post-training included supervised learning, multi-task reinforcement learning and on-policy self-distillation (OPSD) for style alignment. The company also added new reinforcement learning aimed specifically at search ability and safety.
SB Intuitions launched a free beta of the Sarashina service on Oct 13, 2026, according to the post.
API details
The official API docs list the model ID as sarashina3.2-mini-thinking, with a 128k-token context window. It's the only thinking model on the API right now, and the docs say setting reasoning_effort on a non-thinking model returns a 400 error. Output is capped at 8,000 completion tokens per request (max_completion_tokens, default 8000). The docs also list tool calling (function calling) and structured outputs for the Chat Completions API.
Use cases SB Intuitions showcased
The blog walks through four demos:
- Code generation, such as an interactive HTML/JavaScript visualization of torus knots and a Clifford attractor, run in Open WebUI.
- Chat with web search, where the model looks things up and cites sources. In one example, it explains an attack technique from search results but refuses to give reproduction steps.
- Structured data extraction into JSON or Markdown tables, helped by keeping the "thinking" separate from the final formatted output.
- Customer-support agent work with multi-turn tool calls (order lookups and cancellations under a cancellation policy), shown with
reasoning_effortset tolowto keep responses fast.
Benchmarks (company-reported)
SB Intuitions published scores on its docs site comparing the new model with its own Sarashina3 mini. The scores below are on a 0–1 scale (higher is better); the blog notes its own charts use min-max normalization to 0–1. Here are the medium setting results for a few headline benchmarks:
| Benchmark | Sarashina3 mini (no thinking) | Sarashina3.2 mini thinking (medium) |
|---|---|---|
| AIME25 (math) | 0.633 | 0.789 |
| J-tau (Japanese agent tasks) | 0.385 | 0.768 |
| BFCL v3 (function calling) | 0.598 | 0.749 |
| SWE-bench Verified | 0.018 | 0.326 |
| JamC-QA (Japanese culture QA) | 0.821 | 0.694 |
The blog also compares the model with Qwen3-32B, Qwen3-235B-A22B-Thinking-2507, gpt-oss-120b, GPT-5.4 mini and NII's llm-jp-4.1-33b-thinking. SB Intuitions says it scored highest among those models on J-tau, BFCL v3 and the JMT-Safety multi-turn safety benchmark.
The company is upfront about weak spots. It says Japanese culture QA dropped compared with Sarashina3 mini (see table). It also says the model trails Qwen3-235B-A22B-Thinking-2507 and GPT-5.4 mini on GPQA-Diamond and J-MultiChallenge, and that it hasn't yet done enough training specifically for coding-agent use.
Why it matters
SB Intuitions frames Sarashina as "domestic" (国産) generative AI that's strong in Japanese. A from-scratch reasoning model with adjustable thinking depth and strong reported tool-calling scores makes it a more agent-ready option, though every number here is self-reported.
What's confirmed / not confirmed
Confirmed (official SB Intuitions blog and Sarashina API docs):
- Sarashina3.2 mini thinking was announced Oct 29, 2026, and added to the Sarashina service.
- Model ID sarashina3.2-mini-thinking, 128k context, reasoning_effort low/medium/high.
- The Sarashina service free beta started Oct 13, 2026.
- The benchmark figures in the table above, as published by SB Intuitions.
Not confirmed: - Independent benchmark results. All scores come from SB Intuitions, mostly using its own evaluation code (FlexEval), with official implementations for J-tau, tau2-bench and BFCL v3 per the blog. - Parameter count, open-weight availability and paid pricing were not stated in the sources we reviewed.
Related
Up: AI
