According to ChainCatcher, Vitalik Buterin shared his localized, private LLM deployment solution up to April 2026. The core goal is to prioritize privacy, security, and autonomy, minimizing the chances of remote models and external services accessing personal data, and reducing the risks of data leaks, model jailbreaking, and malicious content exploitation through local inference, local file storage, and sandbox isolation.
In terms of hardware, they tested a laptop equipped with an NVIDIA 5090 GPU, an AMD Ryzen AI Max Pro 128 GB unified memory device, and DGX Spark, and used Qwen3.5 35B and 122B models for local inference.
Specifically, the 5090 laptop achieves approximately 90 tokens/s in the 35B model, the AMD solution approximately 51 tokens/s, and the DGX Spark approximately 60 tokens/s. Vitalik stated that he prefers to build local AI environments based on high-performance laptops, while using tools such as llama-server, llama-swap, and NixOS to build the overall workflow.
