Sakana AI starts from-scratch pretraining: Sawara Small kicks off in December, weights due spring 2027
SakanaAILabs · x · 2026-11-13
Sakana AI explained to ITmedia why it is now pretraining its own model, Sakana Sawara (placeholder name), first unveiled by chief scientist Takuya Akiba at its October symposium.
- Designed for autonomous agents, serving as the foundation of its products; full control over data and training recipes keeps R&D transparent.
- Sawara Small pretraining starts in December, weights planned for public release in spring 2027; it doubles as a pilot for a much larger Sawara Large targeting hundreds of billions of parameters.
- Rationale: open-weight models are now practically usable, training knowledge is accessible, costs are falling, and agent-style RL makes domain-specialized models beat general ones.
- Builds on prior agent-layer work (ALE-Agent, AB-MCTS) and model-layer efforts (Sakana Namazu post-training, Sakana Fugu orchestration).
More from Companies & People
- Sunil Pai on starting a new chapter: like joining a school mid-year, winter lock-in ahead — threepointone · 2026-11-13
- WSJ: Cloud providers fleeing Nvidia as wafer commitments hit Huang like an avalanche — SumitGup · 2026-11-13
- Hugging Face CEO Delangue vows to keep fighting for open-source AI "harder than ever" — TinfoilTricorn · 2026-11-13
- Internal Nvidia post urging board to weigh employee sentiment on ex-CEO draws 1,600+ comments — AlexTensor · 2026-11-13
- AIE Code Summit wraps 10th edition with three themes: Factories, Type One models, Dreaming bigger — altryne · 2026-11-13
- Economist publishes 'Jensen Huang must resign' op-ed; Pedro Domingos threatens to cancel subscription — pmddomingos · 2026-11-13