TPN (SN65)’s Genesis Will Put 4GB AI Models on a 2GB Diet

Squeezing a language model down to half its size without losing its intelligence is one of the harder open problems in machine learning.

True Performance Network (TPN), is live on Bittensor mainnet (SN65) with a competition built specifically around that challenge.

Miners now have roughly a week to compress a real open-weight model into a fraction of its original memory footprint, with rankings decided purely by benchmark performance.

The winning submission won’t just take the prize; it becomes the network’s first published proof that the approach works.

The Target And The Challenge

TPN Genesis centers on a single, well-defined optimization problem:

1. The base model is Qwen3.5-2B, a compact open-weight model from Alibaba’s Qwen family, running around 4GB at full precision.

2. The compression target is a 2GB memory ceiling, cutting the model’s footprint roughly in half.

3. The constraint is retained capability, submissions need to preserve what the model knows and how well it reasons, not just shrink the file size.

How Submissions Get Scored

Every entry runs through a fixed evaluation split rather than subjective review:

1. MMLU covers general knowledge and counts for 70% of the score, testing broad factual and reasoning ability across many subject areas.

2. HellaSwag covers common sense reasoning and counts for the remaining 30%, testing whether the compressed model still handles everyday inference correctly.

3. The top five miners split emissions on a 40/25/20/10/5 basis, rewarding not just the winner but the four closest competitors behind it.

Timeline And What Happens Next

The submission window is open, running for roughly 6.5 days through block 8,913,992, followed by about three days of scoring across blocks 8,914,002 through 8,935,592.

The winning model will be published on TPN’s HuggingFace page, with the project’s own portal set to serve as a second home for it once it launches.

TPN Genesis is described as the opening chapter for the subnet rather than a one-off event, with more competitions, updates, and reveals expected to follow.

➛ Click here to view the competition’s dashboard, and here to contribute as a miner or validator