Vitalik Buterin, one of the founders of Ethereum, stated that recent developments in artificial intelligence can directly support Ethereum’s privacy infrastructure. The newest step emphasized by Buterin is that the 2-bit quantized version of the DeepSeek V4 model can run on users’ personal hardware with 90 GB of VRAM space.
Native AI usage and hardware differences
Buterin stated that the ability of the renewed low-memory version of DeepSeek V4 to run without the need for central servers is an important advantage. In this way, individual users can access artificial intelligence in a more accessible way. However, performance varies significantly depending on the hardware used. For example, on Apple hardware, the model can produce 35 tokens per second, but on AMD-based systems, this value is only around 7 tokens.
“Providing proper support for multiple hardware manufacturers makes a significant difference between decentralized AI and AI with a truly open ecosystem,” Buterin said.
Buterin also stated that a tool called LuceBox Hub can run intensive artificial intelligence models faster and more efficiently. In the comparison made using its own RTX 5090 graphics card, it was stated that this tool runs approximately twice as fast as llama.cpp, one of the most common local models. It was also emphasized that LuceBox Hub is still under development.
Mini dictionary: 2-bit quantized model: A compressed version in which weights are represented by only two bits to reduce the size and memory requirement of large language models. In this way, faster results can be obtained with less hardware.
| Hardware | Tokens Per Second (DeepSeek V4 2-bit) | Memory Need |
|---|---|---|
| Apple (MAC) | 35 | 90GB VRAM |
| AMD | 7 | 90GB VRAM |
| RTX 5090 (LuceBox Hub) | about 2x llama.cpp | Variable |
Ethereum-specific AI and privacy infrastructure
Buterin pointed out that native AI infrastructure directly supports Ethereum’s privacy goals. In particular, zero-knowledge proofs (ZK) can enable both paid remote major language model calls and private RPC read operations on the Ethereum network. Thanks to this technology, the technical infrastructure between the two areas is similar and the developments benefit each other.
Mini dictionary: Zero-knowledge proofs (ZK): It is a cryptographic method in which a party proves the accuracy of a certain information without disclosing that information to the other party. It is frequently used in Ethereum’s privacy protocols.
Buterin also stated that artificial intelligence models specifically fine-tuned for the Ethereum ecosystem can increase smart contract and protocol security. For example, the model called Leanstral achieves a speed of 38 tokens/sec in generating Lean code and can compete with models with a much larger parameter. Similarly, trained models for Ethereum-based applications can speed up code verification and security processes.
Buterin’s call for safe protocol
Buterin emphasized the importance of such fine-tuned models in the development process of Ethereum and stated that more systematic and automation-based approaches should be developed in auditing smart contracts. In this context, he pointed out that both artificial intelligence and blockchain technology can work together.
