Skip to main content

Inference subnets

Subnets whose miners serve model output on demand — text, embeddings, classification or scoring — and whose validators grade that output rather than the hardware behind it.

Inference on Bittensor today

17 subnets are classified under inference. 13 publish at least one first-party interface and 17 have a surface that answered our most recent probe. The thing to compare here is not raw capacity but what the validator actually rewards: a subnet scoring answer quality and one scoring latency produce very different miner behaviour, and the API you end up calling reflects whichever it is. Check whether a specification exists before you plan an integration around it.

Bittensor inference subnets, ranked by integration readiness
SubnetReadinessSurfacesHealth
SN2 DSperse1003/24
SN6 Numinous1006/27
SN14 Cacheon1004/25
SN53 engy1005/16
SN59 Babelbit1003/25
SN38 ChronoLLM964/25
SN64 Chutes9619/76
SN96 Verathos967/69
SN110 Green Compute964/110
SN28 gm932/21
SN46 Instant933/51
SN10 Pareton863/12
SN29 hoτfloaτ866/15
SN92 Enclave860/13
SN27 Orion780/11
SN95 Actual740/12
SN31 rec4ll100/5

What "inference" means here

Subnets whose miners serve model output on demand — text, embeddings, classification or scoring — and whose validators grade that output rather than the hardware behind it.

How the classification is derived

Categories come from what a subnet publishes about itself — its declared purpose, its source repository and the interfaces it exposes — not from a hand-maintained list. A subnet can belong to several, and 60 of 129 currently belong to none, which is a coverage gap in our data rather than a statement about those subnets.