Ethereum co-founder Vitalik Buterin says his local AI setup has taken a clear step forward with DeepSeek V4. He highlighted a new 2-bit quantized version that runs within about 90 GB of VRAM and can handle larger workloads on a single high-end machine. He pointed to the Hugging Face build as the current way he runs the model locally.
Vitalik reported that this configuration is “only fast on Apple hardware,” where he has seen around “35 tok/s.” On AMD GPUs, however, he said performance drops to “~7 tok/s,” which makes the same model feel much slower in day-to-day use. He argued that this gap shows why serious local AI must support more than one hardware vendor to serve a broad user base.
He also noted progress on his own messaging-daemon project, which now has alpha Telegram support, even though he still describes the account setup as “quite janky.” In addition, he pointed to Lucebox Hub as a promising way to run “dense” models like Qwen 27B more efficiently. On his 5090 laptop, he said that Lucebox delivers about twice the token throughput of Llama.cpp, even though the software remains experimental.
CROPS AI, Hardware Choice, and Ethereum
Vitalik drew a line between basic “decentralized AI” and what he calls genuine “CROPS AI.” In his view, “actually taking the effort to properly support more than one hardware manufacturer” is central to that difference, and he hopes the ecosystem “can become better at this” as more local-first tools launch.
He added that projects like VoxTerm, which provide local AI recording without third-party servers, reflect a broader shift toward user-controlled AI. He also said, “There’s actually a lot of intersection between ‘CROPS Ethereum access layer’ and ‘CROPS AI,'” noting that a zero-knowledge system for paid remote LLM calls could also be “just as useful” for private RPC reads on Ethereum.
Vitalik highlighted application-specific finetuned models as another key piece, citing Mistral’s Leanstral, where he gets “~38 tok/s on AMD” using under 70 GB of VRAM. He said, “Things like this are a huge boon for writing more secure code,” and tied that to his recent work on formal verification.
He added that “we should have models fine-tuned for Ethereum-related use cases as well,” so AI can help developers spot flaws in smart contracts and protocol code before they reach the mainnet.
READ MORE: Crypto Crash Today: Why Are Bitcoin and AI Coins Falling?
