TPN v0.2 Changes How Optimized AI Models Are Scored on Bittensor

True Performance Network (TPN) has released TPN v0.2, introducing Fluid Bench as a new way to measure whether compressed AI models preserve intelligence. The upgrade will launch with Competition 5 on SN65 and changes both how models are benchmarked and how miners are scored.

At the centre of the release is fluid_knowledge, a benchmark designed to measure relative intelligence instead of relying on fixed public tests. TPN developed it after four competitions showed how difficult it is to judge model quality when miners can study public benchmarks in advance. Fluid Bench generates secret questions from predefined knowledge areas using frontier models, publishes their hashes for verification, then reveals the questions later before generating a new set.

This makes it harder for miners to optimize specifically for known benchmark questions. TPN first scores the original base model, then compares each miner’s compressed version against it to see how much knowledge was lost. Answers are graded from 1 to 10, allowing the system to distinguish between partial understanding and complete loss of knowledge. This is not done to determine which model is smartest overall, but whether optimisation has made a model meaningfully less capable.

TPN v0.2 also simplifies validation. Miners now run the competition benchmark suite themselves through TPN Bench and submit the resulting run IDs with their models. Validators check those records against the TPN backend to confirm the model, scores and authenticity. Valid matches are accepted without repeating the entire benchmark process, reducing duplicated work and speeding up evaluation.

Competition 5 will be the first full test of the new system and will score submissions entirely through fluid_knowledge, removing public benchmarks miners could prepare for ahead of time. TPN noted that the mission remains producing highly optimised open source models that can run on smaller hardware while preserving as much intelligence as possible. Fluid Bench is designed to make the results behind those models easier for miners, validators and users to verify and trust.