A 1.6B decoder-only LM whose training can be independently verified, with midtrained and SFT variants. Record: https://open1b.gensyn.ai
AI & ML interests
We network together the core resource for machine intelligence to flourish alongside human intelligence https://www.gensyn.ai/research
Recent Activity
Papers
The Role of Feedback Alignment in Self-Distillation
IR3DE: A Linear Router for Large Language Models
-
Training-Free Dynamic Upcycling of Expert Language Models
Paper • 2603.29765 • Published • 10 -
Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing
Paper • 2509.08721 • Published • 664 -
All is Not Lost: LLM Recovery without Checkpoints
Paper • 2506.15461 • Published • 41 -
NoLoCo: No-all-reduce Low Communication Training Method for Large Models
Paper • 2506.10911 • Published • 9
A 1.6B decoder-only LM whose training can be independently verified, with midtrained and SFT variants. Record: https://open1b.gensyn.ai
-
Training-Free Dynamic Upcycling of Expert Language Models
Paper • 2603.29765 • Published • 10 -
Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing
Paper • 2509.08721 • Published • 664 -
All is Not Lost: LLM Recovery without Checkpoints
Paper • 2506.15461 • Published • 41 -
NoLoCo: No-all-reduce Low Communication Training Method for Large Models
Paper • 2506.10911 • Published • 9
RL Swarm is an open source system for peer-to-peer gossip-based reinforcement learning over the internet.