Ethereum co-founder Vitalik Buterin has outlined a plan for deploying localized, private large language models, with privacy and security at the center of the design. The proposal aims to reduce the risk of data leakage and unauthorized access by avoiding dependence on remote models and external services.
Local inference, on-device storage, and sandboxing
The approach focuses on three main pillars: local inference, on-device file storage, and sandbox isolation. Instead of sending prompts, files, or context to cloud-based systems, local execution keeps sensitive workloads on the user’s own machine. On-device storage further limits unnecessary data exposure, while sandboxing is intended to constrain model access and create a more controlled runtime environment.
Hardware tests across multiple systems
Buterin also shared performance tests across several hardware setups, including a laptop equipped with an NVIDIA 5090 GPU, a device powered by AMD Ryzen AI Max Pro with 128 GB of unified memory, and the DGX Spark. Running the Qwen3.5 35B model, the 5090 laptop delivered 90 tokens per second, while the AMD system reached 51 tokens per second and the DGX Spark achieved 60 tokens per second.
Preference for high-performance laptops
Based on these results, Buterin appears to favor high-performance laptops as a practical foundation for local AI environments. Such systems offer a balance between performance, portability, and direct user control. He also referenced tools including llama-server, llama-swap, and NixOS, indicating that his vision extends beyond raw compute to model serving, environment management, and system-level reproducibility.
Overall, the plan highlights a broader shift in AI deployment thinking: performance remains important, but privacy, device autonomy, and security isolation are increasingly central. For developers and power users concerned about sensitive data handling, private local LLMs may become an increasingly compelling option.