Can Your Laptop Run Vitalik Buterin’s Local AI?

Ethereum co-founder Vitalik Buterin ran Alibaba’s Qwen3.8‑Flash‑Next model on a laptop built around AMD’s Strix Halo chip and posted timing results on Sept. 17, 2026. The posted table shows the model processed requests entirely on the device without contacting a cloud server. Short prompts returned at a pace comfortable for reading; prompts that reached tens of thousands of words produced slower output.

The laptop uses AMD’s Strix Halo, a single chip that combines central processing and graphics functions and shares a single pool of memory. Shared memory matters because a model must fit into available RAM before it can run. Strix Halo laptops can be configured with up to 128 gigabytes available to the chip.

Typical discrete graphics cards provide around 8 to 24 gigabytes of memory, which is insufficient for many large models. The shared memory in Strix Halo machines allows larger models to run on a single laptop rather than requiring server-class hardware.

Alibaba published the Qwen3.8‑Flash‑Next weights on Aug. 26, 2026. The company describes the model as having 125 billion parameters but designed to activate roughly six billion parameters at a time, a technique that reduces peak memory demand. Buterin used the downloadable Qwen3.8‑Flash‑Next build and reported prompt timing columns for pre-existing prompt, new prompt, generated output, input tokens per second and output tokens per second. He also noted improvements in tools such as llama.cpp that have sped processing on consumer hardware.

Buterin highlighted a privacy benefit of local inference: a model that runs on a device does not transmit the raw query to a cloud provider. For heavier or more specialized tasks, he proposed a hybrid workflow where the local model strips names, wallet addresses and private code before forwarding a sanitized request to a hosted system. He wrote, “Use your local model to orchestrate queries to powerful models so your queries don’t leak your personal information.” He added that such screening would reduce the amount of sensitive data leaving the device but would not guarantee complete protection.

A class action filed in May accuses a major provider of sharing ChatGPT user queries with other companies. Cloud services continue to host the most capable models and the largest inference fleets. The release of open model weights and wider availability of hardware with large shared memory make it easier to run many workflows locally.

Local setups like the one Buterin used can handle many everyday prompts at usable speeds while keeping sensitive inputs on-device. Very large documents and the most demanding inference tasks still favor cloud-hosted systems.