Buterin Tests Local AI on Laptop, Sees Privacy and Speed Gains

The Ethereum co-founder, Vitalik Buterin, believes that local AI is nearly capable of managing a significant portion of daily activities. He tested Alibaba's Qwen3.8-Flash-Next on his personal laptop and shared the performance data.

In contrast to ChatGPT, this configuration does not communicate with any cloud server. The model resides on the device, and the device handles the request independently.

Local AI Test by Buterin Demonstrates Practical Speed

Buterin's laptop is equipped with AMD's Strix Halo chip. Typical computers divide tasks between a CPU and a separate GPU, each with its own memory bank. Strix Halo integrates both onto one chip, allowing them to access a shared memory pool.

This architecture is crucial because an AI model must be loaded into memory to function. Standard graphics cards provide only 8 to 24 GB, insufficient for a model of that scale. Strix Halo systems come with up to 128 GB accessible to either component. Consequently, a single laptop can accommodate a model that previously required server-grade equipment.

The speeds he reported are sufficient for typical tasks. Brief queries returned at a pace comfortable for reading. Generation slowed when prompts reached tens of thousands of words, making lengthy documents still a limitation.

Alibaba released the open weights on August 26. According to the team, the model contains 125 billion parameters but activates only six billion at once, keeping memory requirements low.

Buterin referred to it as Qwen3.8-Flash, but the downloadable version from Alibaba is called Qwen3.8-Flash-Next. Its bigger counterpart, Qwen3.8-Max, achieved notable benchmark results in August.

Qwen 3.8 flash is truly impressive, and llama.cpp has been rapidly getting better and better at processing it

columns are: pre-existing prompt, new prompt, generated, input tok/s, output tok/s

This is on my laptop (strix halo). I think we're very close to the point where you… pic.twitter.com/v5Ze4Fv5Wr

— vitalik.eth (@VitalikButerin) September 17, 2026

How Privacy Alters the Equation

Beyond speed, Buterin notes another benefit. A locally run model responds on the device, meaning no service provider gets the request.

For heavier tasks, he suggests a division of labor. The local model would manage what it can, removing sensitive information from anything it forwards to a larger cloud-based system.

“use your local model to orchestrate queries to powerful models so your queries don’t leak your personal information”

In practice, the local model would extract names, wallet addresses, or private code from a prompt, sending only the leftover question. This filtering would reduce data leaving the device. It would not ensure that no sensitive information escapes.

This stance aligns with his past actions. He has raised concerns about surveillance during the EU chat control debate, and cryptocurrency users have advocated for stricter limits on agents for comparable reasons.

In May, a class action lawsuit alleged that OpenAI shared ChatGPT user queries with Meta and Google.

Cloud providers still dominate the frontier. However, each improvement in local performance shifts more everyday tasks away from their servers, and affordable shared-memory hardware continues to proliferate.

The remaining question is how much performance individuals are willing to sacrifice for control.