[Local Large Models Should Be Designed Around Standardized VRAM Tiers, Vitalik Suggests 96GB as the Upper Limit] Ethereum co-founder Vitalik Buterin stated in a discussion reply that local large model operations should be designed around several standardized "tiers" for memory and VRAM configurations, with 96GB being one of the feasible upper limits for local setups. He mentioned that the RTX 6000 has 96GB of VRAM, high-end MacBooks and some AMD laptops have 128GB of unified memory, but approximately 32GB should be reserved for other tasks. Vitalik suggested that the entire local LLM and GPU ecosystem should standardize hardware and model designs around a few capacity tiers such as 2GB, 8GB, 24GB, 96GB, 256GB, and 1TB to improve deployment efficiency and compatibility.