We've reached a major milestone in fully decentralized training: for the first time, we've demonstrated that a large language model can be split and trained across consumer devices connected over the internet - with no loss in speed or performance.

RT @_AlexanderLong: Thats a wrap for ICML2025. Incredible to watch the space go from "What are you talking about" to "That's impossible" to….

We also made a @3blue1brown style video explaining the paper at a high level.

Pluralis has a main tack paper at ICML this week and the team is in Vancouver running several events. The best will be the Open Source Mixer we're running with @gensynai and @akashnet_ Thursday night. For anyone interested in decentralised training it should be a great evening.

RT @gensynai: Connect with fellow open-source developers, engage with leading minds in decentralized AI (DeAI) and machine learning, and en….

RT @_AlexanderLong: Using beautiful Grafana dashboards for everything internally, so much nicer than Tensorboard. Wandb still good but does….

@SameeraRamasin1 Now this is feasible, we will begin a live run of the first community Protocol Model - Pluralis’s Genesis Run - in the coming days. Registration here: EOT.

@SameeraRamasin1 We also expect this to also have a major impact on centralized training, reducing the importance of high speed networking, making spot-instance based training feasible, and increasing inference speeds.

@SameeraRamasin1 And it is the most significant milestone in our Protocol Learning Research Program.

@SameeraRamasin1 This unlocks:. 1. Truly community-created and owned base models. 2. A new path to scaling base models beyond anything we have seen to date. It is a result that we set out to achieve almost exactly one year ago today.

@SameeraRamasin1 What this means is, for the first time, individuals can pool globally distributed compute to train models far larger than they could alone. As the model is split across nodes, the only constraint is how much compute the protocol can gather.

This is a very high-level summary. All details are in the pre-print (led by @SameeraRamasin1)

By decomposing the high-rank embedding information, we can enforce a shared subspace across all blocks. We then only need to transmit a small set of coefficients between nodes.

We found this property in EVERY transformer we analysed, across all parameter sizes and architectures. It’s obscured by the recursive addition of positional embeddings via residuals - but once found, it can be used to dramatically decrease communication overhead.

How is this possible? All transformer models - regardless of size or architecture - have a hidden property: the output projection weights of each block have low stable rank.

Protocol Models fix this: we place one transformer block on each device and compress the communication between blocks by over 100x without altering training dynamics. This enables multi-participant training, of very large models, over the internet at datacenter speeds.

Why? Because the internet is 100–300× slower than datacenter connections. Compressing to compensate breaks the model's internal communication. Errors build up, training collapses.

What has been achieved? Today, training happens in datacenters, where models are spread over many GPUs with fast interconnects. Trying to replicate this over the internet causes huge slowdowns.

This work proves a third path is viable between closed models and today’s open-weight releases, which remain centralized and unsustainable. Community-owned models are the only true open-source AI and open a new path to scale.

Almost exactly 1 year ago today Pluralis set out to solve a very difficult problem. We are pleased to announce an update on that problem. We can now combine small consumer devices, connected via the internet, to train giant models. Full paper release in 72 hours.