Native nodes
Apple Silicon
M1, M2, M3, M4 family. The CPU's integrated GPU is enough for models up to ~8B in 4-bit quantisation. Apple Silicon shares system memory with the GPU — an M3 Pro with 18 GB of RAM can comfortably run Hermes 8B.
NVIDIA / Linux
Any CUDA-capable GPU works. Larger VRAM and memory bandwidth bump you into higher-tier models, which earn higher reward weights. A single 3090/4090 (24 GB) can serve any model in the catalogue.
CPU-only
Functional for testing, but inference will be slow (10–30× slower than GPU). Reward per task is still calculated, but throughput is too low to be economical.
Browser nodes
Chrome, Brave, Edge, Arc — anything Chromium-based with WebGPU. Hardware shouldn't need anything special; Apple Silicon Macs and modern Windows laptops with 8GB+ RAM handle 1B–3B models comfortably. Mobile browsers technically work but throughput is poor; treat them as experimental.
Model tier guidance
- 1B–2B class → laptops, MacBook Air, mid-range Android phones (browser only)
- 3B class → MacBook Air M-series, mid laptops, ≥8 GB unified or ≥6 GB VRAM
- 7B–8B class → MacBook Pro / Studio, gaming PCs with 12+ GB VRAM
See the model catalogue and per-model reward weights at /docs/models.