Quasar Models launches AI training on Bittensor's 896-expert network

Quasar Models on Tuesday launched a decentralized AI training marketplace built on the Bittensor network, aiming to cut the cost of developing machine-learning models by matching compute buyers with distributed GPU providers.

"Decentralized AI training can reduce costs by as much as 60 percent compared with centralized cloud providers, while also giving smaller teams access to compute they could not otherwise afford," said the company in its launch announcement.

The marketplace operates as a Bittensor subnet, one of 896 specialized expert networks within the system. Bittensor's architecture activates only 16 of its 896 experts per token, routing training tasks to the most efficient providers. Quasar's platform handles model distribution, gradient synchronization, and checkpointing across a decentralized pool of GPUs, removing the need for developers to negotiate directly with individual compute suppliers.

The launch signals growing real-world utility for blockchain-based compute marketplaces, a sector that has struggled to gain traction against centralized providers such as Amazon Web Services and Microsoft Azure. Bittensor's market capitalization stood at roughly $3.2 billion as of Tuesday, according to CoinGecko data, with its TAO token changing hands near $195. Analysts have identified $220 as a key resistance level for TAO, contingent on continued subnet adoption and trading volume.

How the marketplace works

Quasar's platform abstracts the complexity of distributed training. Developers submit a model architecture and training dataset, and the subnet splits the workload across available GPUs, aggregates gradients, and returns a trained model. The system uses Bittensor's consensus mechanism to verify that providers complete their assigned computations correctly, addressing a long-standing trust problem in decentralized compute markets.

The approach competes directly with centralized AI training services from AWS, Google Cloud, and Azure, where a single training run for a large language model can cost millions of dollars. By tapping underutilized consumer and data-center GPUs, Quasar's marketplace could lower the barrier to entry for startups and academic researchers.

Bittensor's expanding ecosystem

Bittensor has attracted a growing number of subnet projects since its inception. The network's competitive subnet model rewards projects based on performance and adoption, creating a merit-based structure that aligns incentives between compute providers and model developers. Several startups building on Bittensor subnets have been accepted into Nvidia's Inception startup program, according to the company.

The Quasar launch follows the release of Moonshot AI's Kimi K3 model, a 2.8 trillion-parameter system that runs on Bittensor's architecture and was ranked first in Frontend Code benchmarks ahead of Anthropic's Claude Fable 5. The model's open-weight release on July 27 gave developers access to the largest publicly available AI system, further validating Bittensor's infrastructure for large-scale AI workloads.

What's at stake

Decentralized AI training faces an uphill battle against the scale and reliability of centralized cloud providers. AWS, Azure, and Google Cloud together control more than 60 percent of the cloud computing market, and their AI-specific services have grown rapidly. Quasar's success depends on attracting enough GPU providers to offer competitive pricing and reliability, while convincing developers that decentralized training can match the performance of dedicated clusters.

If the model gains traction, it could accelerate a shift toward open, community-driven AI development that challenges the dominance of big technology companies. Bittensor's subnet structure allows multiple specialized marketplaces to coexist, each competing on price and performance for different types of AI workloads.

This article is for informational purposes only and does not constitute investment advice.