LightLX
A tool for running Hugging Face language models on Apple Silicon that are too large to fit in RAM, by streaming their weights layer by layer from disk, including large mixture-of-experts models.
LightLX runs Hugging Face models on Apple Silicon that do not fit in unified memory, by streaming weights from disk instead of loading them fully into RAM. It auto-selects between a fast resident mode when a model does fit in memory, loading fully to the GPU like LM Studio or Ollama, a generic streaming mode for any mlx-lm-supported dense or mixture-of-experts model too large for RAM, routing only the active top-k experts per token for MoE models, and a bespoke streaming engine built specifically for the full, unquantized 753-billion-parameter GLM-5.2 model.
Streaming reuses mlx-lm's exact forward pass, so its output is verified bit-identical to loading the model natively. Because streaming throughput is bound by disk read speed rather than compute, it is inherently slow, tens of seconds per token for very large models on typical drives, which the project describes as a hardware limit rather than a bug.
Usage is a single lightlx command that asks for a local model path and opens a chat session.