Vitalik Buterin Tests Local AI on Laptop, Claims It Can Replace ChatGPT

Ethereum co-founder Vitalik Buterin has stated that regional artificial intelligence (AI) has advanced enough to handle many daily tasks. He ran Alibaba's Qwen3.8-Flash-Next model on his laptop, sharing speed results.

Differing from ChatGPT, this environment communicates with no cloud servers at all; the model is installed directly on the device and processes requests immediately.

Buterin's local AI test proves practical speed. His laptop features an AMD Strix Halo chip, which integrates CPU and GPU into a single chip sharing one memory pool—unlike most computers where these components work separately with individual memory allocations.

This architecture is significant because AI models must load entirely into memory before execution. Standard GPUs offer only 8–24GB of memory, insufficient for large models. The Strix Halo-equipped device can utilize up to 128GB of memory, enabling large models previously requiring server hardware to run on a single laptop.

Speeds were sufficient for everyday tasks: short queries received readable responses quickly, though very long questions exceeding tens of thousands of words slowed output. Thus, extremely long documents remain a weakness.

Alibaba released the open weights for this model on August 26, stating it contains 125 billion parameters but only utilizes 6 billion at once, explaining its relatively low memory demand. Buterin referred to it as Qwen3.8-Flash, while Alibaba released a downloadable version under Qwen3.8-Flash-Next. The higher-tier Qwen3.8-Max achieved strong benchmark scores in August.

Buterin highlights a second advantage beyond raw speed: local models respond directly on the device without sending request data to providers. For more complex tasks, he proposes splitting processing—local model handles what it can, removes sensitive information like names or wallet addresses before passing remaining queries to external systems.

This aligns with Buterin's previous stance; he has warned about surveillance in EU chat regulation debates and crypto users have called for stronger AI agent regulations. A May lawsuit alleged OpenAI shared ChatGPT user queries with Meta and Google.

Cloud providers still lead, but improved local performance is shifting daily workloads off servers, with affordable shared-memory hardware becoming increasingly common. How much functionality people will sacrifice for personal control remains unclear.