CrowdGPT Wants to Train an Open-Source LLM on Crowdsourced Home GPUs
Vxtzq1 · reddit · 2026-11-08
A developer is building CrowdGPT, an attempt at the first fully open-source, 100% free LLM trained without datacenters: users contribute their own GPUs in exchange for ownership of the model they train. The author admits it's essentially rebranded federated learning, but with a central server coordinating all participants for a more stable, safer training ground. Training is fully open, data transparent, non-profit; code is on GitHub with a call for feedback.
Related event: CrowdGPT Crowdsources GPU Power to Train First Datacenter-Free 1B Open LLM(3 posts)→
More from Research
- Zyphra's ZHA sparse attention keeps 35B model quality while reading just 1-6% of KV-cache — burny_tech · 2026-11-08
- CoRL 2026 sells out: record 2,094 submissions, 687 accepted papers — Majumdar_Ani · 2026-11-08
- RoboColosseum: 5,699 blind comparisons benchmark physical AI agents vs policies — DJiafei · 2026-11-08
- MuJoCo 4.0 ships with speculative contacts, permanent soft-body deformation and MuJoCo Studio — ZoubinGhahrama1 · 2026-11-08
- mitsuhiko: djs' JS "heap dumps" may beat real dumps for durable coding agents — mitsuhiko · 2026-11-08
- Salesforce paper: RL-trained experts merged via OPD save 42% FLOPs vs external-teacher distillation — rohanpaul_ai · 2026-11-08