Making Large AI Models More Accessible to Home PCs

For more than three months, I have been quietly researching methods to reduce the hardware and memory load required to run very large AI models -particularly mixture-of-experts models such as GLM-5.2-on ordinary home computers.
This work forms part of my master’s degree in computer science, and the early results are promising.
I now have a working prototype, and I plan to share more about it soon.
The objective is not simply to create a smaller version of the model or claim that hundreds of billions of parameters can magically fit inside a consumer GPU.
The complete model remains available, but the system attempts to load, retain and transfer only the components required for the current stage of inference.
My research involves areas such as:
Dynamic expert residency
Predictive expert prefetching
Hierarchical loading across VRAM, system RAM and NVMe storage
Cache-aware routing
Reducing unnecessary parameter movement
Adapting the execution path to the hardware available
I recently saw another project exploring a similar direction, which encouraged me to make my own work public.
However, I believe some current approaches may be underestimating the real inference load.
Counting only the parameters assigned to active experts does not represent the complete cost of inference. Shared layers, attention states, the KV cache, routing decisions, expert transitions, memory bandwidth, page faults and CPU-to-GPU synchronisation can all become major bottlenecks.
A system may appear efficient when measuring only active parameters while still performing poorly during real end-to-end inference because it repeatedly transfers data between storage, RAM and VRAM.
My approach is therefore not focused only on selecting fewer experts.
It also considers where model components should reside, when they should be moved, what should remain cached and how upcoming requirements may be predicted without loading unnecessary parts of the model.
The research is still ongoing, and there is much more testing to complete. However, the results so far suggest that there may be a practical path towards running much larger models on consumer hardware with significantly lower peak memory pressure.
The prototype is already operational, although it is still experimental and requires further optimisation, validation and testing across different hardware configurations.
I intend to share the prototype, or an early public demonstration of it, soon.
The experiments are producing promising results.
And I believe large-model inference on home computers can become considerably more efficient than it is today.
More will be shared soon. #AI