Quasar & Its 10 Trillion Token Run Is Trending on Hugging Face

SILX AI published Quasar-Preview to Hugging Face, and within days, the model appeared on page two of the trending list, alongside Xiaomi, Qwen, and Liquid AI. Independent developers had already built MLX and GGUF ports, so the model runs on MacBooks and local setups without specialized hardware. Nobody coordinated that. They built the conversions because the architecture was worth the effort.

SILX is now matching that community signal with the largest token-scale training commitment in Bittensor's history: a 10-trillion-token decentralized run on Subnet 24, starting immediately.

Below is what Quasar-Preview is, what the 10T-token run targets, and what SN24 miners are being asked to do.

The Architecture Behind the Model

Quasar-Preview is an approximately 18B total parameter Mixture-of-Experts (MoE) checkpoint, with roughly 2B parameters active per inference pass. MoE routing means each query activates only a fraction of the full parameter count, delivering meaningful capability without running all 18B parameters on every token. The model carries an experimental 5M-token context configuration, though under 1B tokens of long-context extension training have been completed so far. SILX AI is explicit: mature 5M-token reasoning quality isn't there yet.

Quasar-Preview is a foundation checkpoint, not a finished model. SILX released it to make the architecture public, give miners and researchers something real to work with, and begin the next phase of decentralized scaling on SN24. The community treated it as exactly that.

What Trending on Hugging Face Actually Means

Every serious open-source model lives on Hugging Face. Page two of its trending list puts Quasar-Preview in front of ML researchers, enterprise developers, and independent builders scanning for capable models. Its current page neighbors are Xiaomi's MiMo series, Qwen's latest releases, and Liquid AI's foundation models.

The MLX and GGUF ports tell a more specific story. When a model releases, it ships in a base format that requires GPU infrastructure to run. The community converts it when it's worth the effort. GGUF is a compressed, portable format that lets you run a model locally through tools like Ollama on any laptop or desktop, no cloud account required. MLX is Apple's framework optimized for Apple Silicon, letting Macs run inference at speeds that used to require dedicated GPU servers. When independent developers build both formats, they're extending the model's reach from data centers to MacBooks, to anyone running a local AI stack.

Nobody at SILX organized those conversions. The community built them before any outreach because Quasar-Preview's architecture was worth deploying. A model people want available in their local setup is a model they intend to work with, not just read about. That's the difference between a release that generates attention and one that generates adoption.

A 10T-Token Run in Two Phases

Prime Intellect's INTELLECT-1, the first globally distributed training of a 10B parameter model, completed its run on 1 trillion tokens. Quasar's planned run targets 10 trillion, ten times that count.

The run has two phases. The first targets 5T tokens to produce a stronger base checkpoint. The second extends the run by another 5T tokens, reaching 10T total trained tokens. SN24 sets the direction for both: the starting checkpoint, the dataset, the training recipe, and the evaluation system. Miners contribute the compute layer. They help Quasar train faster and improve continuously across each checkpoint, earning TAO for doing so.

SILX's core argument is that more useful training matters more than bigger parameter counts. Data quality, training direction, and continuous checkpoint improvement are what close the gap to centralized models. The 10T-token target is the mechanism for testing that argument at scale.

Mining is live. The tracking dashboard is at training.silxinc.com and the subnet repository is publicly available on GitHub.

What SN24 Stakers Are Watching

Most subnets compete on inference quality, speed, or cost. Quasar competes on training progress: checkpoint by checkpoint, verifiable on-chain.

Quasar-Preview is the first public proof of the architecture at real scale. The 10T-token run is where that architecture either narrows the gap to centralized models or defines where decentralized training hits its current ceiling. For SN24 stakers, the subnet's value isn't tied to a roadmap claim. It attaches to a training record visible in real time.

The open-source community already decided Quasar-Preview was worth their effort. The 10T-token run is where the network answers the same question.

Disclaimer: This article is for informational purposes only and does not constitute financial, investment, or trading advice. The information provided should not be interpreted as an endorsement of any digital asset, security, or investment strategy. Readers should conduct their own research and consult with a licensed financial professional before making any investment decisions. The publisher and its contributors are not responsible for any losses that may arise from reliance on the information presented.