The release of Edge0-145B, an open-source 145-billion-parameter model that runs on a consumer laptop, quietly challenges an assumption that has shaped the entire AI industry: that frontier capabilities require cloud infrastructure.
How it works
Edge0-145B uses a latent Mixture-of-Experts architecture, activating only about 10 billion parameters per step while keeping the full model on an SSD. In laptop tests, it consumed 29 GB of memory, processed prompts at over 1,100 tokens per second, and drew roughly 120 watts — comparable to a gaming laptop under load. The model is available for Mac, Windows, Android, and iPhone.
This is not a compressed or quantised version of a larger model. It was trained natively at 4-bit precision, meaning the efficiency is architectural, not retrofitted.
Why it matters
The dominant narrative has been that useful AI requires big cloud bills, corporate gatekeepers, and data that flows through someone else’s servers. Edge0-145B offers a practical alternative: real coding, reasoning, and agent capabilities without per-token fees or internet dependency. The code and weights are on GitHub — anyone can run it.
This matters most for developers and users in parts of the world where cloud access is expensive, unreliable, or politically controlled. It also matters for privacy-sensitive applications that cannot send data to third-party APIs.
The Nvidia turmoil and the rising scrutiny of AI infrastructure concentration make the timing noteworthy. A model that runs on local hardware is one answer to the question of who controls compute. The answer, increasingly, is you.