Vitalik Buterin's Local AI Push: Can Your Laptop Replace ChatGPT?

Ethereum co-founder Vitalik Buterin says local artificial intelligence (AI) is close to handling a large share of everyday tasks. He ran Alibaba's Qwen3.8-Flash-Next on his own laptop and posted the speed results.

Unlike ChatGPT, that setup never contacts a cloud server. The model sits on the machine, and the machine answers the request by itself.

Vitalik Buterin's Local AI Test Shows Usable Speed

His laptop uses AMD's Strix Halo chip. Most computers split the work between a processor and a separate graphics card, and each one keeps its own pool of memory. Strix Halo puts both on a single piece of silicon and lets them share one pool instead.

That design matters because an AI model has to fit into memory before it can run at all. A typical graphics card offers 8 to 24 gigabytes, far too little for a model of this size. Strix Halo machines ship with as much as 128 gigabytes that either half of the chip can use. One laptop can therefore hold a model that until recently needed server hardware.

The speeds he posted are quick enough for ordinary work. Short prompts came back at a comfortable reading pace. Output slowed once a prompt ran to tens of thousands of words, so very long documents remain the weak spot.

Alibaba published the open weights on August 26. The team says the model holds 125 billion parameters yet activates only six billion at a time, which keeps memory demands modest.

Buterin named it Qwen3.8-Flash, though Alibaba ships the downloadable version as Qwen3.8-Flash-Next. Its larger sibling, Qwen3.8-Max, drew strong benchmark scores 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

Why Privacy Changes the Calculation

Buterin sees a second payoff beyond raw speed. A local model answers on the device, so no provider ever receives the request.

For more demanding work, he proposes a split. The local model would handle what it can, then strip the sensitive details out of anything it passes to a larger hosted 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 pull names, wallet addresses or private code out of a prompt, then pass on only the remaining question. Such screening would cut what leaves the device. It would not guarantee that nothing sensitive slips through.

That pitch matches his record. He has warned about surveillance during the EU chat control fight, and crypto users have pushed for tighter limits on agents for similar reasons.

A class action filed in May accuses OpenAI of sharing ChatGPT user queries with Meta and Google.

Cloud providers still own the frontier. Yet every gain in local performance moves more routine work off their servers, and cheap shared-memory hardware keeps spreading.

The open question is how much capability people will trade for control.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.