With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time
True Performance Network (TPN) is Bittensor Subnet 65, a distributed AI model compression and optimisation engine. Its goal is to let people run any AI model on the hardware they already have, by rewarding miners for producing quantized versions of models that are small and still accurate.
How competitions work
Each competition defines a base model, a set of benchmarks with minimum scores and a ranking mode:
Miner flow
Verification
A leader validator checks that each benchmark run really tested the committed model, ranks miners, and checks the top candidates for RAM usage, file integrity and base-model provenance before emissions are distributed. Follower validators reproduce the same weights from the leader’s published results.
True Performance Network (TPN) is Bittensor Subnet 65, a distributed AI model compression and optimisation engine. Its goal is to let people run any AI model on the hardware they already have, by rewarding miners for producing quantized versions of models that are small and still accurate.
How competitions work
Each competition defines a base model, a set of benchmarks with minimum scores and a ranking mode:
Miner flow
Verification
A leader validator checks that each benchmark run really tested the committed model, ranks miners, and checks the top candidates for RAM usage, file integrity and base-model provenance before emissions are distributed. Follower validators reproduce the same weights from the leader’s published results.
What exists today
TPN runs recurring competitions on mainnet, each moving through block-based phases: open, revealing, scoring, distributing and complete. The open-source true-performance-network repository provides:
Product
The website, trueperformancenetwork.com, presents TPN’s mission: making AI models run efficiently on everyday hardware through community-built compression and optimisation.
What exists today
TPN runs recurring competitions on mainnet, each moving through block-based phases: open, revealing, scoring, distributing and complete. The open-source true-performance-network repository provides:
Product
The website, trueperformancenetwork.com, presents TPN’s mission: making AI models run efficiently on everyday hardware through community-built compression and optimisation.
Team
True Performance Network is developed by Taofu (taofu-labs), the same team that previously operated Subnet 65 as TAO Private Network. Public development on the true-performance-network repository is led by the GitHub contributors vstam1 and thearvee.
Community
The team posts on X at @TPN_Labs and runs a Discord server. Contact: [email protected].
Team
True Performance Network is developed by Taofu (taofu-labs), the same team that previously operated Subnet 65 as TAO Private Network. Public development on the true-performance-network repository is led by the GitHub contributors vstam1 and thearvee.
Community
The team posts on X at @TPN_Labs and runs a Discord server. Contact: [email protected].
TPN has refocused Subnet 65 on model compression and optimisation.
TPN has refocused Subnet 65 on model compression and optimisation.
A special thanks to Mark Jeffrey for his amazing Hash Rate series! In this series, he provides valuable insights into Bittensor Subnets and the world of decentralized AI. Be sure to check out the full series on his YouTube channel for more expert analysis and deep dives.
Published on 12 November 2026, this episode sees Mark Jeffrey talk with Mentor from TPN (True Performance Network, Subnet 65) about model compression. Mentor explains that an open-source AI model has to fit into a device’s memory, and that TPN runs miner competitions to find the best way to shrink a model to a target, such as an iPhone or a particular laptop, or to compress it as far as possible while keeping its behaviour. He walks through quantization, for example mapping 16-bit weights down to 8-bit or 4-bit, and distillation, where the original model acts as a teacher for the smaller one, and why intelligence drops off faster the further a model is squashed. Scoring is split between Fluid Bench, a benchmark that regenerates its tests each time so miners cannot simply train on known questions, and KL divergence, which checks how closely the compressed model’s next-word predictions match the original. He shows Gratis AI (gratis-ai.com), which runs compressed models directly in the browser with no app or cloud connection, and describes two kinds of customers: anyone who wants a model quantized for restricted hardware as a service, and inference providers who want to serve more users on the same hardware. The conversation also covers why Bittensor suits this work, including access to talent around the world and miners who constantly look for better approaches, and the balance between protocol changes and long-term stability. Mentor closes by noting that the current competition is testing a much larger Gemma model, that the team’s workbench is public, and that a series of demo apps built on TPN-compressed models is on the way.