AI Inference

QFC v2.0 introduces Proof of Contribution (PoC) compute mode, where validators can contribute AI inference workloads instead of (or in addition to) traditional PoW hashing. Miners run approved ML models locally and submit inference proofs to earn contribution scores.

Want to start mining?

See the Inference Miner Guide for step-by-step setup instructions — works on Intel Mac, Apple Silicon, and Linux.

Compute Modes

Each validator operates in one of three modes:

Mode	Description
pow	Traditional Blake3 proof-of-work hashing
inference	AI model inference tasks
none	No compute contribution (block production only)

Validators can switch modes via the --compute-mode flag:

# PoW mining (default)
qfc-node --validator <KEY> --mine --compute-mode pow

# AI inference
qfc-node --validator <KEY> --mine --compute-mode inference

# Auto-detect: use inference if a supported model can be loaded, otherwise fall back to PoW
qfc-node --validator <KEY> --mine --compute-mode auto

Approved Models

Models must be approved by on-chain governance before inference proofs are accepted. The following models are approved in v2.0:

qfc-embed-small

Property	Value
HuggingFace Model	sentence-transformers/all-MiniLM-L6-v2
Task	Text embedding (sentence similarity, semantic search)
Architecture	BERT (6 layers, 384 hidden dim)
Output	384-dimensional float32 vector
Size	~80 MB
Min Memory	512 MB
GPU Tier	Cold (CPU-only nodes)

qfc-embed-medium

Property	Value
HuggingFace Model	sentence-transformers/all-MiniLM-L12-v2
Task	Text embedding (higher quality than small)
Architecture	BERT (12 layers, 384 hidden dim)
Output	384-dimensional float32 vector
Size	~120 MB
Min Memory	512 MB
GPU Tier	Cold (CPU-only nodes)

qfc-classify-small

Property	Value
HuggingFace Model	google-bert/bert-base-uncased
Task	Text classification
Architecture	BERT (12 layers, 768 hidden dim)
Output	768-dimensional float32 vector
Size	~440 MB
Min Memory	2048 MB
GPU Tier	Warm (GPU recommended)

GPU Tiers

Nodes are classified into tiers based on available hardware:

Tier	Hardware	Models Available
Cold	CPU only	qfc-embed-small, qfc-embed-medium
Warm	Entry GPU (4-8 GB VRAM)	All Cold models + qfc-classify-small
Hot	High-end GPU (16+ GB VRAM)	All models

How Inference Mining Works

1. Model Download — On startup, the miner downloads approved model weights from HuggingFace Hub and caches them locally in ~/.cache/qfc-models/.

2. Task Assignment — Each epoch, the AI coordinator assigns inference tasks to registered miners based on their GPU tier and loaded models.

3. Inference Execution — The miner runs the model locally (CPU or GPU via candle) and produces an inference proof containing the input hash, output hash, and execution time.

4. Proof Submission — Proofs are submitted via the qfc_submitInferenceProof RPC method and broadcast to the network via P2P.

5. Spot-Check Verification — 5% of proofs are randomly re-executed by the receiving node. If the result doesn't match, the validator is penalized (5% stake slash + 6-hour jail).

6. Score Update — Valid inference proofs update the validator's inferenceScore in the PoC contribution calculation, which weights 20% of the total score.

Configuration

Environment Variables

Variable	Default	Description
QFC_COMPUTE_MODE	pow	Compute mode: pow, inference, or auto
QFC_INFERENCE_BACKEND	auto	Backend: cpu, metal, cuda, or auto
QFC_MODEL_DIR	/models	Directory for cached model weights

Docker Example

# docker-compose.yml
services:
  qfc-inference-miner:
    image: ghcr.io/qfc-network/qfc-core:latest
    environment:
      QFC_COMPUTE_MODE: inference
      QFC_INFERENCE_BACKEND: auto
      QFC_MODEL_DIR: /models
    volumes:
      - model-cache:/models

RPC Methods

qfc_submitInferenceProof

Submit an inference proof for scoring.

curl -X POST http://localhost:8545 \
  -H "Content-Type: application/json" \
  -d '{
    "jsonrpc": "2.0",
    "method": "qfc_submitInferenceProof",
    "params": [{
      "validator": "0x...",
      "model_id": "qfc-embed-small@v1.0",
      "input_hash": "0x...",
      "output_hash": "0x...",
      "execution_time_ms": 106,
      "flops": 1000000000,
      "signature": "0x..."
    }],
    "id": 1
  }'

qfc_getValidators

Returns validator list including compute mode and inference scores.

{
  "address": "0x8d1dd4...",
  "computeMode": "pow",
  "providesCompute": true,
  "hashrate": "5629542",
  "inferenceScore": "0",
  "tasksCompleted": "0"
}