FreeToken serves a 753B open-weight model from a single workstation GPU

A Berkeley-led team (Shuo Yang, Xiaoze Fan, Melissa Pan and others, with Kurt Keutzer, Song Han, Matei Zaharia and Ion Stoica) posted FreeToken to arXiv on 17 August 2026: a serving system that treats a personal machine as one elastic inference platform and maps mixture-of-experts computation to whatever bandwidth is actually available. The paper reports a 35B model running on an 8GB laptop GPU, a 284B model on a gaming desktop card, and the 753B GLM-5.2 on a single workstation GPU, across more than 20 MoE models and real coding and tool-using agents.

What it means The gap between "open weights exist" and "I can run them" is a bandwidth-scheduling problem, and this is the clearest published attempt to close it on consumer hardware.