A reported experiment ran a two-bit-quantized GLM-5.3 744B model on a high-end workstation. Model files alone occupied about 238.6 GB, using two 96 GB GPUs together with system RAM. The author measured generation at 11.3 tokens per second, but a 32,984-token input took about 394 seconds to produce the first response. Lowering reasoning intensity allowed a simple game implementation to finish, while the high-intensity setting exhausted the budget. This is one example on equipment costing around $65,000, not evidence of cost efficiency on ordinary PCs or quality equivalent to the original model.