Zorlo: Running Large Mixture-of-Experts Models on Low-RAM PCs by Streaming Experts from the SSD

Mixture-of-Experts (MoE) language models are large on disk but use only a small part of their weights for any one token. Qwen3-30B-A3B stores about 30 billion parameters and activates about 3 billion per token. Zorlo is an inference engine and desktop app for Windows that uses this sparsity to run such models on ordinary low-RAM PCs with no GPU. The weights that every token needs stay in RAM (0.82 GB for Qwen3-30B-A3B). The experts stay on the SSD and are read on demand into a cache that sizes itself to the memory the PC has free. On a 4-core laptop with 8 GB of RAM, Zorlo decodes Qwen3-30B-A3B at about 3.4 tokens per second with a 1.2 to 1.5 GB expert cache, and at 2.3 to 3.0 with a cache of only 256 MB. llama.cpp reaches 1.43 tokens per second on the same machine, reading the same file through memory mapping, and Zorlo's default output is token-for-token identical to llama.cpp's. Optional speed modes reach 5.5 tokens per second with a measured quality cost. The same engine runs the 117-billion-parameter GPT-OSS-120B at 1.5 to 1.7 tokens per second on that laptop, where llama.cpp manages 0.37. This report describes the design at a high level, the measurements behind these numbers, and the current limits.