wtf is xav-glm52-2box-local?

This repository shows how to run GLM-5.2, a very large AI model with 753 billion parameters, entirely on two Apple Mac computers with no internet connection. The model is so large that a single Mac with 128GB of memory cannot hold it, resulting in a speed of under one token per second. By connecting two Macs together with a single cable into 256GB of shared memory, the full model stays in memory and runs at about 18.5 tokens per second. The repository provides patches and scripts to reproduce this setup. One patch fixes a software crash that happens when large models communicate over a network between the two machines. Another patch adds support for a specific part of the model architecture. The launcher script checks the connection between the two computers, confirms enough memory is free, runs a test, and starts the model. It cleans up memory if anything fails. The recommended configuration uses a setting called f16 for memory and reduces the number of active experts from eight to five. This increased speed by nearly nineteen percent over the starting setup without changing the model or affecting output quality. The setup served up to 128,000 tokens of context during live testing. The model itself is compressed to about 202GB using a method called IQ1_S. The repository includes demo tools used to create proof videos. One demo asks the model to write a complete Pac-Man game in a single file. The resulting file is included and works in a browser with one character fixed. Another demo streams text generation in real time with a live speed counter, and a script renders the output into a video. The project builds on existing open source work from llama.cpp, Z.ai, and Unsloth. The model weights are open.