Setup Guide

Three steps to start mining modelOS and earning MDL + PRL.

Get a wallet address

Create one in the modelOS desktop wallet or online at compute.modeloslab.xyz. Use your MDL address for block rewards. Optionally add a PRL address to also earn Pearl via AuxPoW — the pool handles upstream Pearl submission and distributes both currencies via the same PPLNS window.

Pick how you run

Requires an NVIDIA GPU (Ampere SM80 / RTX 3070 or newer; Blackwell & Hopper get the highest throughput). Choose your platform below — Docker on your own rig, a rented GPU on RunPod / Vast.ai, or HiveOS. Paste your address and copy the config for that platform.

Run and start earning

Launch it. Stats appear within minutes of your first share. Check earnings at pool.modeloslab.xyz/lookup.

Configuration

Start mining

MDL wallet address

Required — block rewards paid here

PRL wallet address (optional)

Add your Pearl address to also earn PRL via AuxPoW — the pool handles the rest

Worker name (optional)

Own NVIDIA GPU? One command — needs Docker + the NVIDIA Container Toolkit.

docker run -d --restart unless-stopped --gpus all --shm-size 8g \
  -e POOL_HOST=stratum.modeloslab.xyz \
  -e POOL_PORT=5566 \
  -e POOL_TLS=false \
  -e MDL_ADDRESS=mdl1p… \
  -e POOL_WORKER=rig-01 \
  -e HF_TOKEN=hf_… \
  -e MODELOS_MODEL=8b \
  -e MODELOS_MODEL_VERSION=4 \
  -e MODELOS_TENSOR_PARALLEL_SIZE=1 \
  -e MODELOS_GPU_INDICES=0 \
  -e VLLM_PORT=8000 \
  -e MINER_NO_GATEWAY=true \
  -e MINER_NO_VLLM_PLUGIN=true \
  -e MODELOS_GPU_MEMORY_UTILIZATION=0.60 \
  -e MODELOS_MAX_MODEL_LEN=8192 \
  -e MODELOS_COORDINATOR_URL=https://pool.modeloslab.xyz \
  -e MODELOS_WORKER_PRIVKEY=<worker-private-key> \
  -e MODELOS_DIRECT_MINER_V2=1 \
  -v modelos-models:/model-cache \
  modeloslab/vllm-miner:latest

stratum.modeloslab.xyz:5566 (PPLNS) · stratum.modeloslab.xyz:5567 (SOLO) · 2% fee

VRAM: inference + mining needs room for the model and the ~6 GB the proof reserves (fixed — it can't be reduced). A 32 GB RTX 5090 is the lowest GPU that fits a model alongside the miner — it runs the 8B tier. 24 GB (4090) and 8 GB (3070) cards must use Miner only.

8B	~16–18 GB	RTX 5090 32 GB (lowest), A100, H100
14B	~28 GB	A100 40/80 GB, H100, 2×24 GB (TP=2)
32B	~64 GB	A100 80 GB, H100, multi-GPU (TP)
70B	~140 GB	2×80 GB A100/H100 (TP=2)

Tune MODELOS_GPU_MEMORY_UTILIZATION: 0.60 is conservative — it leaves the miner ~6 GB on a 32 GB card. Raise iton larger cards (A100/H100) or multi-GPU rigs to fit bigger models; lower it if vLLM OOMs. For multi-GPU, set MODELOS_TENSOR_PARALLEL_SIZE to the number of cards the model spans.

Optional — Earn More

Also earn inference fees

On a GPU with enough VRAM (≈32 GB+), drop MODELOS_MINER_ONLY=1 and add a model + HF_TOKEN — the same vllm-miner image then serves AI inference while it mines. Every confirmed inference_tx triggers a race; the first valid proof wins the locked fee, paid directly to your MDL address with no pool cut.