Ethereum co-founder Vitalik Buterin has highlighted that recent advances in artificial intelligence could directly enhance Ethereum’s privacy infrastructure. Buterin’s latest focus is on the DeepSeek V4 model’s 2-bit quantized version, which can now run on personal hardware with 90 GB of VRAM, potentially making powerful AI locally accessible to more users.
Running AI locally and hardware differences
Buterin pointed out that the new low-memory release of DeepSeek V4 is notable because it no longer depends on centralized servers, which is a major step forward. This enables individual users to interact with state-of-the-art AI directly from their own computers. However, performance varies significantly depending on the hardware. For example, Apple systems can generate up to 35 tokens per second with the model, whereas AMD-based systems reach only about 7 tokens per second.
“Delivering proper support for multiple hardware manufacturers is what makes the real difference between genuinely open AI ecosystems and merely decentralized ones,” Buterin emphasized.
Additionally, Buterin mentioned LuceBox Hub, a tool designed to run heavy AI models faster and more efficiently. In his tests using an RTX 5090 GPU, LuceBox Hub outperformed one of the most popular local solutions, llama.cpp, by almost two times. He added that LuceBox Hub remains in active development.
Glossary: 2-bit quantized model: This is a compressed version of a large language model where the model weights are represented using only two bits, which reduces memory requirements and allows faster performance on lower-end hardware.
| Hardware | Tokens per second (DeepSeek V4 2-bit) | Memory needed |
|---|---|---|
| Apple (MAC) | 35 | 90 GB VRAM |
| AMD | 7 | 90 GB VRAM |
| RTX 5090 (LuceBox Hub) | about 2x llama.cpp | Varies |
AI designed for Ethereum and privacy
Buterin noted that local AI infrastructure can directly support Ethereum’s privacy goals. For instance, zero-knowledge proofs (ZK) can power both paid remote large language model calls and private RPC reads on the Ethereum network, thanks to the shared technical foundations of these two fields. Advancements in one area can thus benefit the other.
Glossary: Zero-knowledge proofs (ZK): A cryptographic method that allows one party to prove the validity of certain information to another party without revealing the actual information. Used widely in Ethereum’s privacy protocols.
Buterin also stated that AI models fine-tuned specifically for the Ethereum ecosystem could boost smart contract and protocol security. As an example, he referenced the Leanstral model, which can generate Lean code at a rate of 38 tokens per second and is competitive with much larger models. Similar Ethereum-specific models could speed up code verification and security audits across decentralized applications.
Buterin’s call for secure protocol development
Highlighting the key role of such fine-tuned models in Ethereum’s evolution, Buterin called for more systematic, automated approaches to smart contract audits. He underlined that artificial intelligence and blockchain should be developed to work together for enhanced security and reliability.
