For advanced
Self-hosting: step-by-step guide
At Gonka, GPUs process network neural network requests and receive GNK tokens for each completed request.
Don't want to deal with servers?
You don't need your own equipment to earn GNK. Join a pool — invest from $30 and get GNK without technical skills.
Hardware Requirements
Network Node (network server)
- 16-core CPU (amd64)
- 64GB+ RAM
- 1TB NVMe SSD
- 100Mbps+ network
ML Node (compute node)
- MLNode 3.0.12+ minimum
- Go 1.22.8, Docker Desktop 4.37+, Java 19+
- NVIDIA GPU (newer generation than Tesla)
- Minimum 40GB VRAM per MLNode container
- For current network MoE FP8 models: ~640GB VRAM total (320GB per node), minimum 2 ML-nodes
- Reference configuration: 8xH200 per node
- CUDA 12.6-12.9, NVIDIA Container Toolkit
Tip: No GPUs of your own? Rent bare-metal from GPU providers from ~$11,600/month for 8xH100.
Gonka.Top — the best referral program in the ecosystem
Collective plan from $100. 25% L1 + 5% L2 of pool profits. Monthly payouts for 180 days.
Setup Steps
01 Install inferenced CLI
Download the latest version (v0.2.10+) from GitHub releases.
# Download and install inferenced CLI
wget https://github.com/gonka-ai/gonka/releases/latest/download/inferenced-linux-amd64
chmod +x inferenced-linux-amd64
sudo mv inferenced-linux-amd64 /usr/local/bin/inferenced02 Create a cold wallet key
The cold wallet is used to manage the node and receive rewards.
inferenced keys add cold-walletSave your mnemonic phrase in a safe place!
03 Clone the repository and configure
git clone https://github.com/gonka-ai/gonka.git
cd gonka
cp config.env.example config.env
# Edit config.env — enter your parameters04 Download model weights
Active network models — MoE FP8 (currently MiniMax M2.7 and Kimi K2.6; the set is defined by governance). Weight loading via HuggingFace CLI. подробнее →
pip install huggingface-cli
# актуальный id модели — из конфига сети / governance (пример: MiniMax M2.7)
huggingface-cli download MiniMaxAI/MiniMax-M2.705 Start Docker services
docker compose pull
docker compose up -d06 Register on-chain
Create an ML operational key and register the node on the network.
# Create operational key
inferenced keys add ml-operator
# Register node
inferenced tx register-ml-node --from cold-wallet
# Grant operator permissions
inferenced tx grant-ml-permissions --from cold-wallet --to ml-operator07 Configure SSL and launch
Configure DNS (Cloudflare/AWS/GCP/Azure/DO/Hetzner), obtain an SSL certificate, launch the full network, and verify registration.
# Check node status
inferenced status
# Verify registration
inferenced query node-infoDon't want to set it up yourself?
There are GPU hosting and pools that do everything for you:
Dedicated server
- Gonka.Top Dedicated — Gonka.Top — ~$12,000/month, full turnkey service, 100% of the power is yours
Mining pools
- Gonka.Top — from $100, 25% referral + 5% L2, monthly payouts
- GonkaPool.ai — Telegram bot, convenient entry
- CloudMine (Mingles) — contracts by "weight" 50–10000, payment USDC on Base
- Gonka.Wallet — Telegram Mini App, one-click wallet, built-in pools
Best referral: Gonka.Top — 25% of profits. Lowest entry point: Ancapex — from $1, daily payouts.
What Affects Profitability
- Weight: base weight = 20% automatically. The remaining 80% is activated by staking GNK as collateral. подробнее →
- GPU performance: the more powerful the graphics card, the more neural network requests are processed per round.
- Network utilization: at 40-60% load the price is stable. Above 60% — price increases (more revenue per request).
- Fair rewards: The system automatically verifies each computation (<10ms) and fairly distributes rewards. During network updates, there is a transition period (~5 hours) so that no one loses income. подробнее →