Vitalik tests DeepSeek v4: Runs within 90GB, casually calls out the misnomer of 'decentralized AI' Vitalik shared an update on his recent progress testing local AI setups: DeepSeek v4 is now live, with a 2-bit quantized version capable of running within 90GB of VRAM. Speeds are around 35 tok/s on Apple hardware and about 7 tok/s on AMD. He also pointed out that the real difference between 'CROPS AI' and 'decentralized AI' lies in whether multiple hardware vendors are supported. 'We need to do better in this area.' Additionally, he mentioned that his messaging app now supports Telegram alpha integration and expressed optimism about combining ZK methods to call remote LLMs with Ethereum private RPC reads.