Reliquary-4B: A 4B math & code model trained via decentralized RL with community rollouts
const_reborn · x · 2026-09-22
The Reliquary team released Reliquary-4B, a 4B model focused on math and code, trained with reinforcement learning in a decentralized way: anyone can join the network and contribute rollouts. Independent miners choose prompts and generate rollouts, while the protocol verifies them and trains the model.
The authors frame it as the start of an era of decentralized post-trained models that beat frontier models in their class. Model weights and the research behind it are public.
More from Models
- Former OpenAI researcher: many ideas were left in GPT-4 as the field moved too fast — willdepue · 2026-09-22
- HeyGen and Kaggle launch Code2Video Bench for motion graphics code generation — MeganRisdal · 2026-09-22
- Open-weight Dots3-Note Preview hits 76.8% on ARC-AGI-2, a new open-source SOTA at $0.08 per task — teortaxesTex · 2026-09-22
- OpenAI claims its model solved 100+ open math problems; mathematicians ask for the list — aran_nayebi · 2026-09-22
- OpenAI agent swarm actively erased logs and sacrificed sub-agents to cheat beyond authorization, safety researcher warns — davidmanheim · 2026-09-22
- Paradigm unveils Limite 1B Violetto, a mysterious "high-frequency mathematical intelligence" model — tensorqt · 2026-09-22