I need to rant about local models

The creator begins by outlining the gap between "runnable" and "good" local AI models, attributing this to several factors: the high cost of hardware, challenges with parallelism, and prohibitive electricity costs. They contrast the capabilities of locally run models with cloud-based solutions, noting that while GLM 5.2 is impressive, it's not a model that runs well on consumer hardware. The speaker then delves into the specifics of hardware costs, citing the exorbitant prices of high-end GPUs like the RTX 5090 and RTX 6000 Ada Generation, and suggests that for those wanting to run powerful models locally, investing in multiple consumer GPUs or a professional workstation might be necessary. The creator also touches on the efficiency metrics and pricing of various LLM providers on OpenRouter, comparing models like GPT-5.5 and GLM 5.2. The video concludes by listing several problems with the current state of local LLMs, including the performance gap, hardware costs, parallelism issues, and electricity expenses, ultimately suggesting that cloud solutions might be more practical for many users.