Chutes AI Trained Model on Distributed Consumer GPUs
The startup successfully trained an 8 billion parameter model using hardware scattered across 13 countries.
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 demonstrated a new decentralized training system called Parallax, which successfully trained an 8 billion parameter model using 240 consumer RTX 5090 GPUs. The hardware was distributed across 30 hosts in 13 countries at a total cost of $6,500.
Why it matters
The system aims to reduce the high barriers to entry for large-scale AI development by enabling decentralized training on consumer-grade hardware. This approach could significantly lower costs for training large models compared to traditional, centralized data centers.
The 8 billion parameter model achieved 59.6 tokens per second on mobile CPUs, while a larger 40.75 billion parameter variant maintained 26.9 tokens per second using 12GB of peak memory.
The players
Jon Durbin
He is the researcher who presented the Parallax decentralized training system at the Exploit Summit.
Chutes AI
This is an AI startup and Bittensor subnet SN64 that focuses on developing decentralized machine learning training solutions.
The details
Presented at the Exploit Summit in Montreal, the Parallax system stitches together hardware from around the globe using libp2p for machine synchronization. Chutes AI, which operates as Bittensor subnet SN64, plans to release both a technical report and the model publicly.
Timeline
Jon Durbin presented the Parallax system at the Exploit Summit held on September 28-29, 2026.
The Tech Race
This development advances the capabilities of the Bittensor protocol by demonstrating that large-scale model training can be performed within its decentralized subnet architecture. It signals a move away from reliance on expensive, centralized data centers toward a model of global, crowd-sourced compute resources.
Developers and researchers may soon gain access to cheaper, decentralized training methods, lowering the cost of creating advanced AI models. Additionally, the ability for these models to run efficiently on mobile CPUs suggests future smartphone applications will support more complex local AI functionality.
The takeaway
The success of Parallax suggests that the future of large-scale AI training may not be restricted to those who can afford massive, private server clusters. Decentralized compute networks are proving that consumer hardware, when properly networked, can compete with traditional industrial infrastructure.
Further reading
For more information on the evolving landscape of model development, explore our Artificial Intelligence section.
Source note: This article includes information reported by Crypto Briefing.
Live Poll
Do you trust that decentralized AI models running locally will better protect your personal data privacy?
