Decentralized System Trained AI Model on Consumer GPUs
The Parallax system utilized global consumer hardware to train models capable of running on mobile CPUs.
Updated on Sept. 30, 2026 in Artificial Intelligence
Live Poll
Do you trust that decentralized AI models running locally will better protect your personal data privacy?
Chutes AI has successfully trained an 8 billion parameter model using a distributed network of consumer-grade hardware. The system, known as Parallax, demonstrated that AI training can be performed across geographically dispersed infrastructure.
Why it matters
This development marks a shift toward decentralized model training, potentially lowering the barrier to entry for high-performance AI development. By bypassing centralized server farms, it offers a path to utilizing underused consumer hardware at a significantly reduced cost.
The Parallax system utilized 240 RTX 5090 GPUs spread across 30 hosts in 13 countries to achieve training at a cost of $6,500. The resulting 8B model achieves 59.6 tokens per second on mobile CPUs, while a 40.75B parameter variant runs at 26.9 tokens per second using 12GB of peak memory.
The players
Chutes AI
A decentralized AI research project operating as a Bittensor subnet focused on distributed model training.
Jon Durbin
An AI researcher and presenter who introduced the Parallax training system at the Exploit Summit.
The details
Parallax operates as a Bittensor subnet, a decentralized protocol for incentivizing machine learning compute, that stitches together hardware scattered across the globe. It relies on libp2p — a modular network stack for peer-to-peer communication — to synchronize data and weight updates between the distributed machines in real time. This approach effectively treats fragmented consumer GPUs as a unified training cluster for large-scale models.
Timeline
Jon Durbin presented the Parallax system at the Exploit Summit in Montreal on September 28-29, 2026.
The Tech Race
This effort follows a pattern set by the Bittensor decentralized AI protocol, which seeks to challenge the dominance of centralized cloud providers in model training. It demonstrates a competitive trajectory where distributed, low-cost hardware attempts to match the capabilities of large-scale, enterprise-grade GPU clusters.
The ability to run large language models on mobile CPUs at these token speeds suggests future mobile applications could perform complex reasoning tasks without relying on cloud servers. Availability for these models and the full technical documentation are pending upcoming announcements from the Chutes AI team.
The takeaway
The successful training of large models on distributed consumer GPUs signals that the infrastructure barrier for AI development is beginning to lower. Readers should monitor upcoming publications from Chutes AI to access the full model weights and verify the performance benchmarks in their own environments.
Further reading
For more on the latest research in distributed compute, see our coverage of Artificial Intelligence.
Live Poll
Do you trust that decentralized AI models running locally will better protect your personal data privacy?
