Training a competitive AI model typically requires the kind of GPU budget that makes venture capitalists weep. Quasar, operating as Subnet 24 on the Bittensor network, thinks it has a workaround: let a decentralized army of miners do it instead, at a fraction of the cost.

The project, developed by SILX AI under the SILX Labs umbrella, has built an open marketplace where independent miners can operate and refine AI models through a competitive evaluation system. Miners who improve model performance get rewarded. Those who don’t, well, they get outcompeted.

What Quasar is actually building

At its core, Quasar is focused on a specific pain point in AI: long-context foundation models. Quasar is targeting context lengths of approximately 2 million tokens or more. To get there, the team is deploying what it calls Quasar Attention, a novel methodology paired with hybrid architecture approaches designed to push the boundaries of what decentralized training can produce.

The Quasar-3B architecture launched in April 2026, following the release of the Quasar-Preview model as a Mixture-of-Experts checkpoint on Hugging Face. It’s computationally efficient, which matters a lot when your training infrastructure is spread across a decentralized network rather than sitting in a single data center.

The real headline number: Quasar claims a 99.5% reduction in pre-training costs compared to traditional centralized methods.

The 10-trillion-token ambition

Quasar has laid out plans for a 10-trillion-token decentralized training run, split into two phases of 5 trillion tokens each.

For context, GPT-3 was trained on roughly 300 billion tokens. A 10-trillion-token run would put Quasar in the conversation with the largest training efforts ever attempted, except this one would be executed by a distributed network rather than a single corporate entity.

Cross-subnet collaboration

Quasar announced a collaboration with Subnet 56, known as Gradients, and Subnet 3 for integrated long-context model training. This partnership, formed around July 2026, points to an emerging pattern within Bittensor where subnets are starting to function less like isolated experiments and more like interconnected components of a larger system.

The competition-and-reward mechanism is central to how Quasar maintains quality. Miners submit model improvements, validators evaluate them, and rewards flow to the contributors producing genuine enhancements.

What this means for investors

For anyone watching the intersection of AI and crypto, Quasar represents an interesting test case. The TAO token, which powers the broader Bittensor network, stands to benefit if projects like SN24 can demonstrate that decentralized training produces models competitive with centralized alternatives.

Investors should watch two things closely. First, whether the Quasar-3B model produces benchmark results that hold up against comparable centralized models. Second, whether the 10-trillion-token training run actually launches and progresses on schedule.