Overview

What is Pareton?

Models are converging in quality; cost and latency are the real competitive edge. At the same time, inference is becoming the dominant share of AI compute (longer generations, agent loops, and training that depends on high-volume serving) while new model families and accelerators keep expanding the configurations worth evaluating. Kernels, batching, caching, quantization, and scheduling compound into a search space no small team can exhaust by hand.

Pareton continuously discovers, validates, and deploys the optimal serving configuration for a workload, across models, hardware, and serving stacks.

It runs on Bittensor Subnet 10. Contributors submit small, reviewable git patches against a pinned baseline engine. Pareton gates those patches, benches them in rounds against a frozen customer profile, and seats a campaign leader only when a challenger beats the incumbent on the same prompt set.

How it works

1. Customer profile — You define the production setup (model, hardware, serving configs, workload) and the correctness and scoring rules. One success metric locks the goal.
2. Contributor patches — Miners propose focused changes against the current baseline.
3. Validate — Candidates must build and satisfy path and API constraints before they enter a round.
4. Round — The leader and up to 5 challengers run the same sampled prompt set on one GPU pod. Scores are comparable inside that round only. See Rounds and leaders.
5. Seat or hold — A challenger takes the crown only if it beats the incumbent by the overtake moat. Otherwise the leader holds, or the crown is vacated.
6. Repeat — The leader runs again in every later round.
7. Emission — The seated leader earns subnet emission while it holds the crown. Everything else is burned. See Validators and emissions, or the Validator guide to run one yourself.

Links

Status

Build in progress.