PeerCompute Brings Peer-to-Peer Model Hosting and Inference to the Browser
PeerCompute combines libp2p, WebGPU, and IPFS in the browser for model storage, peer inference, persistent agents, and multiplayer worlds. Several GGUF and ONNX models have passed generation tests; full browser transfer of some large models remains unproven.
Browser as Model Host
PeerCompute is a browser-based peer-to-peer compute platform built on libp2p, WebGPU, and JavaScript workers. NodeKernel coordinates NetworkManager transport, StateManager persistence, and ComputeManager-owned JS/WASM/WebGPU workers. Model artifacts are content-addressed with IPFS/UnixFS, verified and retained in IndexedDB. A browser can retain model files, serve them to peers, and run inference for another player without that player downloading the model. The Models window combines a saved-model library with Hugging Face search, filtering, and compatibility inspection; interrupted imports reuse completed files and supported partial-download checkpoints.
Chat, Agents, and Worlds
The chat interface accepts typed text, pasted or dropped images, and speech input via a hold-and-release Talk button. The default Mixed router prefers a capable loaded local model and otherwise uses an available peer; Local only prevents remote dispatch. Sharing stored files and hosting inference are separate controls. The console is an installable PWA with an offline app shell. Persistent general agents store identity, instructions, goals, context, and bounded conversation history in private IndexedDB records. Agents can pause, resume, cancel, export, and delete. CubeChat and EarthGrid reuse the same agent, model, and speech foundations.
Qualified Models
The shipped workers cover Transformers.js ONNX text/vision, wllama/llama.cpp GGUF text, ONNX diffusion, and separate speech specialists. On Chrome155 with an RTX5060Ti, Qwen3-0.6B q4, SmolVLM 256M/500M q4, and multimodal Qwen3.5-2B q4 pass generation, cancellation, and browser restart with external model requests blocked. The exact Tsotchke-Corporation/Qwen3.8-27B-GeoRefine-D14-abl-GGUF is catalogued for storage, sharing, and GGUF text inference. Its CLI import verified and pinned all 11 files (14.20 GB). The 27B model passes text generation, arithmetic, and warm reuse on the RTX5060Ti16GB with 56 GPU layers, 512 context, and 64 batch, loading in 143.3 seconds. Full 14GB browser peer transfer and browser persistence remain unproven. DreamShaper 8 LCM and SD-Turbo pass managed-worker generation; a GPU-disabled browser receives matching 512×512 pixels over encrypted WebRTC without downloading weights or starting a worker. Warm remote requests measured about 1.7 seconds for DreamShaper and 2.1 seconds for SD-Turbo on the test machine.
Limits and Planned Work
Model routing chooses a capable host per request, not splitting one model’s tensors across peers. Learned routing, consented agent migration and cloning, self-directed training, neural locomotion, and evolving 3D bodies remain planned. The November 8, 2026 chat-control release passed 193 focused cases, the complete 15-demo build, and public model-console/CubeChat/EarthGrid/overview checks. Memory figures are planning estimates, not measurements of free VRAM. Browser storage is scoped to the exact origin, including the port.