Question 1

How does Verathos verify that a miner ran the correct model?

Accepted Answer

Each miner produces a cryptographic proof alongside every inference response—either a Merkle-tree-based proof-v2 that commits to layer-by-layer activations, or a ZK-based proof-v3 that binds output tokens to a signed execution profile. Validators replay these proofs independently; a failed verification triggers immediate probation and EMA score halving, with the result published to the shared routing state within the same block.

Question 2

Can I use the Verathos API with an existing OpenAI client library?

Accepted Answer

Yes. The public endpoint at `api.verathos.ai/v1` is fully OpenAI-compatible. Set `base_url` to that address, supply an API key from the account portal, and use `model='auto'` for score-weighted routing or any specific model ID returned by `GET /v1/models`. No other code changes are required.

Question 3

What GPU hardware is required to mine on Subnet 96?

Accepted Answer

Miners need at least 24 GB VRAM (RTX 4090, RTX 3090 Ti, L4, or A10 class) with CUDA 12.8+, 32 GB system RAM, and a 100 Mbps symmetric connection. The recommended configuration is 48 GB VRAM (A6000, A40, or L40S). Multi-GPU machines can register multiple model endpoints, each pinned to its own GPU.

Question 4

What is the GGUF mesh pool and how does it differ from a vLLM endpoint?

Accepted Answer

A GGUF mesh pool distributes a single large model across multiple GPU machines coordinated by a pool manager. Workers dial out to the coordinator so NAT traversal is handled automatically, and each GPU proves its computation share independently. The vLLM path runs the full model on one machine and uses the same cryptographic scoring; the choice depends on model size and hardware availability.

Question 5

How are miner scores calculated and how do proofs affect emissions?

Accepted Answer

Each epoch (~72 minutes), validators pull signed receipts from miners, match them against pre-planned canary obligations, and compute an EMA score weighted by throughput, latency, model utility (parameters, context length, quantization), and proof validity. A single proof failure halves the EMA immediately; repeated failures or missed canary obligations enter the miner into probation, zeroing its weight in Bittensor's emission calculation until it accumulates clean passes.

Question 6

Does Verathos support pay-per-request without an account?

Accepted Answer

Yes. The subnet integrates x402, a USDC pay-per-request protocol built on Base, which allows callers to pay for individual completions without creating an account or holding an API key. See the user guide for the payment flow details.

Verathos

Verathos

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup

Validator Setup