mohit-jpg-dot/airdrop-llm is a pipeline-parallel inference broker that pools the GPUs and memory of multiple computers on a local wireless network to split a single language model inference across them. It uses Apple Silicon's Metal and NVIDIA's CUDA paths side by side, runs fully offline with no dependencies, and bundles an OpenAI-compatible API gateway. The repository was created on October 19, 2026, and has no public releases.
The README describes the repository as a tool for connecting computers on the same local wireless network to split and run language model inference. Device discovery uses UDP, model layers are placed device by device, static weights stay on each device as they are, and only active values (activations — the computation results passed between layers) are transferred over TCP with compression. The structure is one where inference is distributed across devices at the layer level. It is written in Rust and licensed under MIT. It has 7 stars.
The points to check are the README's large-model support and speed claims. Neither has been confirmed by separate measurements. Based on the published repository information and the README, the scores are 7.0 for freshness, 6.0 for technology, 5.0 for completeness and 2.7 for momentum, with only momentum coming out low. Whether different consumer hardware can be temporarily joined via a local hotspot LAN to run a large model in split form is the deciding factor, and for now the only basis for that is the README's description.