> For the complete documentation index, see [llms.txt](https://docs.robincompute.org/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.robincompute.org/providers/native-worker.md). # Native Worker `robincompute-node` is one Rust binary that carries the whole job: it drives your GPU backend, fetches model weights, works the job queue, and talks to the settlement contracts on-chain. Next to the browser tier, native workers serve far larger models and earn more per job. Add an active $RCOMPUTE stake and every completed job pays out at 85% of its credit value. *** ## What you need | Requirement | Details | | --------------- | ---------------------------------------------------------------------------------------- | | OS | Linux (recommended), macOS, Windows | | GPU | NVIDIA (CUDA), Apple Silicon (Metal), or AMD (ROCm) | | VRAM | 8GB at minimum, enough for the smallest native-tier models | | Ethereum wallet | Needed to register and to receive USDG (MetaMask, Rabby, or Robinhood Wallet) | | Disk | 50GB+ free for weights; a 70B model alone takes \~40GB | | Network | A stable link. The closer you sit to users latency-wise, the better for your reputation. | *** ## Installing the daemon ### Linux (CUDA) ```bash curl -sSL https://install.robincompute.org | bash ``` The script fetches `robincompute-node` with the CUDA backend, checking your NVIDIA driver version first to pick the matching binary variant. Confirm the install: ```bash robincompute-node --version # robincompute-node 0.1.0 (cuda-12.4) ``` ### macOS (Apple Silicon) ```bash curl -sSL https://install.robincompute.org/mac | bash ``` On macOS the node runs Metal through llama.cpp; every Apple Silicon generation (M1 through M4) works. The chip's unified memory is addressed directly, so a 36GB M3 Max has plenty of room for 27B parameter models. ### Windows (CUDA) Grab the installer from `robincompute.org/downloads` and run the `.exe`. It puts `robincompute-node` on your PATH and sets up the CUDA backend for you. ### AMD (ROCm) ```bash curl -sSL https://install.robincompute.org/rocm | bash ``` You need ROCm 5.6 or later, on RDNA 2 hardware (RX 6000 series) or newer. *** ## Bringing a node online ### Step 1: Register ```bash robincompute-node register --wallet ``` Registration does three things: * generates a keypair for the node itself, distinct from your wallet * sends a registration transaction to the `worker_registry` contract * has your wallet sign that registration Your wallet asks you to approve the transaction (delivered by QR code or deeplink). The node's keypair lands in `~/.robincompute/worker.json`. Back it up. This key signs every proof-of-completion you submit, and your on-chain reputation is bound to it. ### Step 2: Add models ```bash robincompute-node models list # browse the catalog robincompute-node models add qwen3-8b robincompute-node models add llama-3.3-70b ``` Weights come down from the RobinCompute model registry into `~/.robincompute/models/`. Every download is verified against a SHA-256 checksum before the model joins your serving list. Want the VRAM footprint before committing to a download? Ask for the model info: ```bash robincompute-node models info llama-3.3-70b # Model: llama-3.3-70b # Parameters: 70B # Format: GGUF Q4_K_M # VRAM required: 43GB # Tier: Max # Earnings per job: $0.300 (unstaked) / $0.340 (staked) ``` ### Step 3: Stake $RCOMPUTE (optional, worth doing) You earn without staking. A stake earns you more: the 85% payout rate plus elevated routing priority. The full walkthrough lives in the \[Staking guide]\( ). ### Step 4: Run it ```bash robincompute-node start ``` On startup the node: 1. loads your chosen models into GPU memory 2. advertises its capabilities to the orchestrator mesh 3. starts pulling jobs 4. prints each completion and its earnings to the terminal For unattended operation on Linux, install it as a service: ```bash robincompute-node install-service systemctl enable robincompute-node systemctl start robincompute-node ``` *** ## Configuration Everything lives in `~/.robincompute/config.toml`. ```toml [node] wallet = "0x7c41f9b8d2a6e3054cf18a9b62d47e0c93f5a1b8" worker_keypair = "~/.robincompute/worker.json" [models] # Models loaded when the node starts active = ["qwen3-8b", "mistral-7b"] [gpu] # Detected from your hardware automatically; set explicitly to override. # Valid values: cuda, metal, rocm backend = "cuda" # Upper bound on the share of VRAM robincompute-node may claim vram_limit = 0.85 [networking] # Listening port for inbound orchestrator connections port = 8765 # Set manually if public IP detection gets it wrong # public_ip = "1.2.3.4" [logging] level = "info" # Also write logs here, in addition to stdout file = "~/.robincompute/logs/node.log" ``` *** ## Watching your node While the node runs, a local dashboard serves at `http://localhost:3001` with: * current jobs and how deep the queue is * hourly earnings and lifetime USDG * GPU utilization and VRAM consumption * your on-chain reputation score * job counts and latency broken down per model The same node exposes a local status API: ```bash curl http://localhost:3001/status | jq ``` ```json { "status": "running", "wallet": "0x7c41f9b8d2a6e3054cf18a9b62d47e0c93f5a1b8", "active_models": ["qwen3-8b", "mistral-7b"], "jobs_completed_24h": 342, "usdg_earned_24h": 27.36, "reputation_score": 847, "stake_active": true, "payout_rate": "85%" } ``` *** ## Upgrading ```bash robincompute-node update ``` This pulls the newest binary and swaps it in place of the old one. Nodes running as a service restart automatically once the swap completes. Model weights survive upgrades. You only re-download if a new model format version ships. *** ## Choosing hardware | GPU | VRAM | Recommended models | Estimated earnings / hour | | -------------------------------- | ----- | ----------------------------------- | ------------------------- | | RTX 3060 | 12GB | Qwen3 8B, Mistral 7B | $0.15 to $0.30 | | RTX 3090 / 4090 | 24GB | Qwen3 8B, Mistral 7B, Llama 3.2 13B | $0.40 to $0.80 | | 2x RTX 3090 (multi-GPU, Q1 2027) | 48GB | Llama 3.3 70B, DeepSeek R1 | $1.50 to $3.00 | | M2 Max 38GPU | 32GB | Qwen3 8B, Llama 3.2 27B | $0.50 to $1.00 | | M3 Ultra 76GPU | 128GB | Llama 3.3 70B, DeepSeek R1 | $1.50 to $3.00 | | A100 80GB | 80GB | All models | $2.00 to $4.00 | These figures assume a steady stream of jobs. What you actually earn moves with network demand, which models you host, and where your reputation score sits. And every payout behind these estimates is a transaction you can check. {: .note } Pooling two or more GPUs under one node is planned for Q1 2027. Until then, each node drives exactly one GPU. --- # Agent Instructions This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com. ## Querying This Documentation If you need additional information that is not directly available in this page, you can query the documentation dynamically by asking a question. Perform an HTTP GET request on the current page URL with the `ask` query parameter, and the optional `goal` query parameter: ``` GET https://docs.robincompute.org/providers/native-worker.md?ask=&goal= ``` `ask` is the immediate question: it should be specific, self-contained, and written in natural language. `goal` is optional and describes the broader end goal you are ultimately trying to accomplish on behalf of the user. GitBook uses it to tailor the answer towards what is most useful for that goal. The response will contain a direct answer to the question and relevant excerpts and sources from the documentation. Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.