At the Exploit Summit in Montreal, Chutes AI introduced Parallax, a decentralized system for training AI models. The project utilizes a distributed network of consumer-grade hardware rather than traditional data centers.
The demonstration featured an 8 billion parameter model trained using 240 RTX 5090 GPUs across 30 hosts in 13 countries. The total training cost amounted to approximately $6,500, or $11 per billion tokens processed.
The resulting 8B model achieves a performance of 59.6 tokens per second on mobile CPUs. A larger 40.75 billion parameter version was also showcased, which operates at 26.9 tokens per second while requiring 12GB of peak memory.
Parallax leverages the libp2p networking framework to synchronize data across global nodes. Chutes AI operates as Bittensor subnet SN64, focusing on serverless decentralized compute for open-source models.
