Sakana AI Starts Pretraining Its Own Model from Scratch, Weights Due Spring 2027
tkasasagi · x · 2026-11-13
Sakana AI announced "Sakana Sawara" (tentative), a foundation model pretrained entirely from scratch and designed for autonomous AI agents, intended as the base of its products and services.
Key points:
- "Sawara Small" starts pretraining in December, with weights planned for open release in spring 2027; it doubles as a pilot for the larger "Sawara Large," targeting models of several hundred billion parameters or more.
- Until now Sakana mainly adapted open-weight foreign models for Japan via post-training (Sakana Namazu, Sakana Fugu), alongside agent research like ALE-Agent and AB-MCTS.
- Rationale: open-weight models are now production-viable, training data and know-how are more accessible, and costs are falling — while agent-based RL can specialize models to beat general-purpose ones.
- Owning the full model and training recipe keeps the entire development process transparent, enabling broader research.
Related event: Sakana AI to Pretrain Its Own Model Sawara from Scratch(4 posts)→
More from Companies & People
- CZ calls Nvidia letting Jensen Huang resign 'a mistake' as post hits 3.1M views — CodeByPoonam · 2026-11-13
- Nvidia employees reportedly call for Jensen Huang's return as CEO in internal thread with 1,600+ comments — Polymarket · 2026-11-13
- Bindu Reddy: Jensen Huang should withdraw his resignation if he committed no crimes — bindureddy · 2026-11-13
- Six PIs at ELLIS Institute Tübingen Named ELLIS Fellows and Scholars — orvieto_antonio · 2026-11-13
- First Jensen-less GTC in March put Ian Buck to the test — zephyr_z9 · 2026-11-13
- DeepMind staff ask UK government to intervene in unionization fight — BlackHC · 2026-11-13