Japan's "domestic LLM" label covers more than one product shape. DigIntoAI maps the scene into three editorial routes—not an official Japanese government or industry taxonomy: (1) API and closed proprietary Japanese models; (2) open-weight Japanese models; and (3) Japan-focused post-training and Japanization, including specialization of overseas open bases and further training on Japanese open weights. The map matters because buyers often mix a government trial roster with later commercial flagships, or treat every Japan-branded API as the same sovereignty story.

Route 1: API and closed proprietary models

This route is the productized stack you call through an API or run under a vendor contract, without open weights as the main offer.

Preferred Networks' PLaMo 3.0 Prime, formally released on June 22, 2026 (Beijing time), is positioned as a from-scratch Japanese foundation model with Reasoning and Non-reasoning variants, API or on-premises options, and a context window extended to 256k. NTT's May 19, 2026 update to tsuzumi 2 stresses a lightweight Japanese business-document model that can read charts and tables in document images and is often used on-premises or in private cloud for sensitive work. SoftBank group company SB Intuitions launched a free beta of Sarashina Service on October 13, 2026, with API access to six Sarashina3-family models plus web chat and translation; the press release says usage is subject to rate limits and that the free beta end date is not yet set. DigIntoAI has not signed up for that beta.

SoftBank's April 16, 2026 press materials describe generative AI services using the homegrown Sarashina LLM on Cloud PF Type A with Oracle Alloy, with staggered availability from June 2026, aimed at data-sovereignty needs for enterprises and local governments. Treat that Alloy / Cloud PF Type A story as SoftBank's sovereignty-oriented cloud path—not as a claim about every Japan-facing API.

Route 2: Open-weight Japanese models

This route is weights you can download and run under a stated license, with Japanese pretraining or strong Japanese focus.

On April 3, 2026, Japan's National Institute of Informatics announced LLM-jp-4 8B and LLM-jp-4 32B-A3B, trained on a high-quality corpus of about 12 trillion tokens and released under an open-source license. Community directories such as LLM-jp's awesome-japanese-llm list also track CyberAgent's CALM3 and Stockmark's Stockmark-2-100B line among scratch or heavily Japanese-trained open releases. Swallow-family models appear in the same ecosystem, but many Swallow entries are continual pretraining (CPT) on overseas bases rather than full indigenous pretraining from scratch—do not drop them into a "built entirely in Japan from zero" bucket without checking the base.

Route 3: Japan-focused post-training

This DigIntoAI route covers extra mid-training or post-training aimed at Japanese business use, whether the starting base is an overseas open model or a Japanese open-weight series.

Sakana Namazu is a Japanese-specialized LLM built on the open model Kimi K2.6 and tuned with Sakana's in-house data for Japanese and Japanese business work. Sakana's pricing page lists no monthly fee: text input $0.95 / 1M tokens, output $4.00 / 1M tokens, cached input $0.15 / 1M tokens, plus tool fees such as web search and code execution billed separately. Yen figures on the page are reference conversions; billing is in US dollars. For DigIntoAI coverage of the Kimi line Namazu starts from, see Kimi K3.1 Has Shown Up in Kimi Code, but There's No Release Yet.

Namazu's FAQ is explicit on residency: at present, Sakana cannot guarantee that data processing is completed entirely within Japan. Do not put Namazu in the same sovereignty tier as SoftBank's Alloy / Cloud PF Type A pitch or the Digital Agency's 源内 government trial.

On October 2, 2026, ELYZA announced ELYZA-Thinking-1.0-llm-jp-4-33b and ELYZA-Thinking-1.0-llm-jp-4-32b-a3b: mid-training and post-training on NII's llm-jp-4 open series, released for commercial use. That is post-training on a Japanese open base, not Japanization of an overseas base—keep the base clear even though DigIntoAI still files the release under the post-training route.

源内 shortlist vs later flagships

Separately from the three product routes, Japan's Digital Agency published on March 6, 2026 the results of a call for domestic LLMs to trial inside Government AI (project name 源内 / Gennai): 15 applications, 7 selected for trial evaluation. The selected models, in the agency's order, are: NTT DATA's tsuzumi 2; Custom Cloud's CC Gov-LLM; the KDDI–ELYZA joint entry Llama-3.1-ELYZA-JP-70B; SoftBank's Sarashina2 mini; NEC's cotomi v3; Fujitsu's Takane 32B; and Preferred Networks' PLaMo 2.0 Prime. That March 2026 roster is not interchangeable with later flagships such as PLaMo 3.0, Sarashina3, or the October 2026 ELYZA-Thinking llm-jp-4 line. Fujitsu's Takane materials also describe collaboration with Cohere—so Takane should not be filed as a zero-base fully indigenous stack either.

How to use this map

Pick the route that matches how you will actually run the model: closed API or contract, self-hosted open weights, or a post-trained specialty API/weights. Then check the dated roster: 源内 for the March 2026 government trial list; vendor release notes for current flagships. For Asia coding-plan and API pricing context on DigIntoAI, see Claude Code with GLM, Kimi, Qwen and MiniMax Plans. DigIntoAI has not independently benchmarked the Japanese models named here and does not claim coding or Japanese-task rankings against Claude, GPT, or other overseas systems.

Sources

  • Digital Agency, Government AI domestic LLM selection results (源内), March 6, 2026: https://www.digital.go.jp/news/10d55c63-b3e1-42b9-9cc5-93a06943ae0e
  • Preferred Networks, PLaMo 3.0 Prime release, June 22, 2026: https://www.preferred.jp/ja/news/pr20260622
  • NTT, tsuzumi 2 update, May 19, 2026: https://group.ntt/jp/newsrelease/2026/05/19/260519a.html
  • SoftBank, Sarashina on Cloud PF Type A with Oracle Alloy, April 16, 2026: https://www.softbank.jp/en/corp/news/press/sbkk/2026/20260416_02/ (Japanese detail: https://www.softbank.jp/corp/news/press/sbkk/2026/20260416_02/)
  • SB Intuitions, Sarashina Service free beta, October 13, 2026: https://www.sbintuitions.co.jp/en/news/press/20261013_01/
  • National Institute of Informatics, LLM-jp-4 release, April 3, 2026: https://www.nii.ac.jp/en/news/release/2026/0403.html
  • Sakana AI, Namazu product page (base model, pricing, FAQ): https://sakana.ai/namazu/
  • ELYZA, RSI Research and ELYZA-Thinking llm-jp-4 releases, October 2, 2026: https://elyza.ai/news/2026/10/02/rsiresearch
  • ITmedia AI+, ELYZA-Thinking llm-jp-4 coverage, October 2, 2026: https://www.itmedia.co.jp/aiplus/article/2610/02/2000001965/
  • Fujitsu, Takane launch (Cohere collaboration): https://www.fujitsu.com/global/about/resources/news/press-releases/2024/0930-01.html
  • LLM-jp, awesome-japanese-llm directory (secondary): https://llm-jp.github.io/awesome-japanese-llm/