Sakana AI to pretrain its own LLM Sawara from scratch, starting December, weights open in 2027
tkasasagi · x · 2026-11-13
Sakana AI detailed its plan to pretrain its own LLM, Sakana Sawara (tentative), first unveiled by chief scientist Takuya Akiba at its October symposium.
- Positioning: designed for autonomous AI agents and as the foundation of its products; full control over data and training recipes for research transparency
- Roadmap: pretraining of "Sawara Small" starts in December, weights to be released in spring 2027; a larger "Sawara Large" targeting hundreds of billions of parameters is planned
- Why now: open-weight models have become practically useful, training knowledge is more accessible, costs are down, and specializing models via agent-style RL end-to-end can beat general-purpose models
- Previously Sakana focused on post-training (Namazu) and model orchestration (Fugu); this marks its entry into the pretraining layer
More from Companies & People
- Ex-Nvidia Employee Says He Was Laid Off, Breaking Jensen's 'No Layoffs' Rule — ruilong_li · 2026-11-13
- FT Reports Nvidia Board Probe Into Huang Affair, Driven by GC Tim Teter — SumitGup · 2026-11-13
- Over 1,600 comments on internal Nvidia post urging board to bring back Jensen Huang — rohanpaul_ai · 2026-11-13
- Texas A&M Builds One of the Strongest US Academic Supercomputers, Hiring 10 — yuz9yuz · 2026-11-13
- Tencent Reportedly DeepSeek's Biggest Backer as China Compute Centralizes — teortaxesTex · 2026-11-13
- kimmonismus urges Logan K to ship an always-on agent beating Meta Muse — kimmonismus · 2026-11-13