a third path
Frontier models trained and owned by the many
Pluralis is a research lab building Protocol Learning: a way for many independent participants to pool ordinary hardware over the internet and train a single foundation model that no one party holds or controls.
Attending NeurIPS in Sydney? Come by our workshop on December 11.
Training runsLive dashboard
- Pluralis-13BSystems test of peer-to-peer routingSep 2026 · 62B tokens
- Pluralis-8B330 contributor nodesMay 2026 · 532B tokens
- Node0-7.5BFirst open run, on Agora’s predecessorSep 2025 · 37B tokens
Recent researchAll papers
- Communication-Efficient LLM Adaptation over Decentralized GPU MeshesNeurIPS 2026
- NuMuon: Nuclear-Norm-Constrained Muon for Compressible LLM TrainingNeurIPS 2026
- AsyncMesh: Fully Asynchronous Optimization for Data and Pipeline ParallelismNeurIPS 2026
- Factored Gossip DiLoCo: Reducing Blocking Communication within DiLoCoICML 2026