Vitalik Buterin, co-founder of Ethereum, has unveiled a detailed guide for running large language models (LLMs) locally, advocating for self-sovereign AI setups. In a recent blog post, Buterin shared his experience using open-weight models like Qwen3.5:35B on an NVIDIA 5090 laptop, achieving speeds of 90 tokens per second. This setup, which avoids reliance on cloud services, emphasizes privacy and control, using NixOS for system configurations and llama-server for local model hosting. Buterin also proposed the development of AI models specifically fine-tuned for Ethereum applications. These models would be trained on Ethereum-specific data, enhancing their ability to understand and reason about the blockchain's unique features, such as its EVM architecture and DeFi protocols. This vision positions Ethereum as a foundational layer for decentralized AI agents, leveraging its existing infrastructure for transactions and identity verification.