Researchers have introduced Edge0, a streaming inference engine designed to mitigate the memory wall encountered when serving large Mixture-of-Experts (MoE) models on consumer-grade hardware. While 35B-class models require significant storage—approximately 19.5GB when quantized to 4-bit—traditional SSD offloading fails to improve performance because the sequential dependency of layers prevents pre-emptive data fetching. To resolve this, Edge0 utilizes a per-layer "prerouter" head that predicts the routing for the subsequent layer one token in advance. By using these predictions as the actual routing instructions, the engine ensures the staged expert set perfectly matches the required routed set, effectively hiding SSD latency. To compensate for the precision loss from 4-bit quantization and the accuracy degradation caused by substituting actual routing with predicted routing, the authors employ an unmerged recovery LoRA trained on the student path. Experimental results demonstrate that Edge0 can serve a 35B MoE model at a throughput of 20 tokens per second on a single 24GB GPU, maintaining a peak active memory footprint of only 3GiB. This performance remains within a few percentage points of the original fp16 teacher model across five public benchmarks. The framework also supports an 8B tier and has been released as open-source software.