Back

Vitalik Buterin Says DeepSeek V4’s 2-Bit Version Advances Local AI on Ethereum Privacy

Ethereum co-founder Vitalik Buterin said DeepSeek V4 can run locally with about 90 GB of VRAM, a development he linked to privacy-focused artificial intelligence use cases.

ETH

Summary

Ethereum co-founder Vitalik Buterin said DeepSeek V4 now has a 2-bit quantized version, referring to a model-compression method that reduces computing requirements, allowing it to run with about 90 GB of VRAM (video memory used by processors). He said the setup reaches roughly 35 tokens per second on Apple hardware. The update was presented as part of his progress on local AI, which refers to running artificial intelligence models directly on user-controlled devices rather than relying on remote cloud infrastructure. In the context of Ethereum privacy, local AI can matter because processing sensitive data on-device may reduce external data exposure.

Terms & Concepts
  • 2-bit quantization: A model-compression technique that reduces numerical precision to lower memory and hardware requirements for running AI systems.
  • VRAM: Video random access memory, the dedicated memory used by graphics processors or similar hardware to handle computation-heavy workloads.
  • Local AI: Artificial intelligence that runs on a user’s own device or hardware instead of a remote server, which can improve data privacy and control.