An open-source project called Edge0 is taking a different approach to running a large mixture-of-experts model on local hardware, Geeky Gadgets reports, serving a Qwen 35B model while keeping the active memory footprint close to 3GB instead of holding every expert in RAM. Edge0 is described in its documentation as a streaming MoE inference framework that keeps selected experts in a hot cache and streams other weights from an SSD when the router predicts they are needed, targeting the memory wall that makes larger local models difficult to run. Its GitHub repository lists two supported checkpoints: edge0-35b, based on Qwen3.6-35B-A3B, and edge0-8b, based on a Ling hybrid model, with the project saying the checkpoints include its trained adapter and router components so the pipeline can be started through the framework’s command line. The documentation says the 35B setup uses quantized experts and staged loading, and describes prefetching and an LRU-style hot stack to limit data movement during generation. An accompanying paper on the memory wall reports a test on a single 24GB machine, with about 20 tokens per second and roughly 3GiB of peak active memory, describing results across five public benchmarks and comparing the local system with a full-precision teacher model. Edge0 remains open-source, and the sources do not establish that it is a drop-in replacement for cloud inference or that the reported speed applies to all