Sakana AI announces LLM pretraining project Sakana Sawara
A startup known for post-training and orchestration moving into full pretraining signals rising confidence that frontier-adjacent training is viable outside the US-China giants. Developers and enterprises evaluating non-US foundation models should track Sawara Small's December start as a potential Japan-built alternative to open-weight options like Qwen or Llama. The stacked project naming (Namazu, Fugu, Sawara) also clarifies how the lab segments its research roadmap.
Sakana AI has revealed plans to begin pretraining its own large language model, provisionally named Sakana Sawara, as featured in an ITmedia AI+ interview with the company. Chief Scientist Takuya Akiba first presented the project, referred to as "Sakana Sawara (tentative name)", at a symposium in October. The company plans to start pretraining a smaller variant, "Sawara Small (tentative name)", in December.
The move builds on the company's earlier work at other layers of the stack: Sakana Namazu focused on post-training, while Sakana Fugu addressed orchestration. Sakana AI's stated view is that the moment is now right to take on pretraining at the model layer itself, completing its coverage of the full pipeline from training to deployment.
Most AI startups in Japan and elsewhere rely on open-weight foundation models rather than training from scratch, so Sakana AI's shift toward pretraining distinguishes it from peers that concentrate on fine-tuning or agent frameworks. The interview outlines the company's background and aims behind this decision.
ITmedia AI+ さまに、Sakana AIがなぜ今、大規模言語モデルの事前学習に取り組むのかを取り上げていただきました。
10月のシンポジウムで弊社Chief Scientistの秋葉拓哉が発表した「Sakana Sawara（仮称）」について、背景やねらいをお答えしました。
Sakana Namazuで取り組んできた事後学習、Sakana Fuguのオーケストレーションに引き続き、今、モデルの層で事前学習に挑む好機が来たと考えています。12月にSawara Small（仮称）の事前学習を開始 する予定です。
記事はこちら🐟️ https://t.co/FJAwLyY8iL
This report is categorized under AITOP's "Models" desk. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Sakana AI, Sakana Sawara, pretraining, open-source models.
Source attribution: curated from Sakana AI (@SakanaAILabs).
