Score (SN44) Delivers Four Production-Ready Models Under 10MB and 80ms

Score (SN44) posted the first results of the latency loop update that tied vision miner rewards to sub-100ms inference performance, with four production-ready models now live on the subnet across crime, fire, carwash, and roadsign detection.

Every model comes in under 10MB with inference latency between 60ms and 80ms, and accuracy ranging from 75% to 90% depending on the task. The numbers are small enough to run on a camera, a gateway, or an embedded board without cloud dependency, which is the exact deployment surface the subnet was built for.

The team’s target is accuracy per byte per millisecond, with miners converging on it faster than expected.

The Four Live Models

The current SN44 leaderboard covers four production-grade tasks, each with a specific footprint the subnet is publishing openly.

The important observation on the table is not any single number but the pattern across all four models: Under 10MB, under 80ms, and production-grade accuracy across four unrelated visual tasks.

That combination is what makes edge deployment on-device possible without stepping back to cloud inference for anything real-time.

Why This Matters

Score’s latency loop tied rewards to inference under 100ms, and the first results confirm the incentive structure did the work it was designed to do. Miners converged on the target faster than the team expected, producing four production-ready models on the leaderboard rather than a single benchmark demonstrator.

The frontier the loop is now optimizing sits at accuracy per byte per millisecond, which is a different objective than what centralized vision AI research targets. Small footprint, low latency, and enough accuracy to be trusted in production is the exact combination that unlocks one small AI brain per camera on the edge, and more model categories are already coming.

➛ Read More on How SN44 Ties Miner Rewards to Sub-100ms Inference Performance: