Researchers have developed a system called FreeToken that enables large language models previously requiring datacenter-scale infrastructure to run locally on personal computers, from laptops with modest GPU memory to gaming desktops and single workstation GPUs.
The system, presented in a paper by Shuo Yang and ten co-authors, addresses a fundamental challenge in the AI field: while frontier-class open-weight models have become widely available, the expectation that serving them requires datacenter infrastructure has limited who can actually use them.
Edge-Native Architecture
FreeToken treats a personal machine not as a small GPU but as a unified, elastic inference platform. The researchers co-designed the full serving stack—including model layout and loading, expert residency management, CPU-GPU execution coordination, agentic state reuse, and runtime memory management—around two realities of local AI deployment.
First, agent workloads continuously change their execution patterns rather than following predictable routines. Second, edge hardware presents heterogeneous resources whose balance varies significantly from machine to machine. Rather than committing to a fixed offloading strategy, FreeToken continuously maps computation and model state onto whatever resources are actually available at runtime.
Hardware Support and Performance
The system supports more than twenty Mixture of Experts models and real coding and tool-using agents across a wide hardware range. On consumer hardware, the researchers demonstrated serving a 35-billion parameter model on a laptop, a 284-billion parameter model on a gaming desktop, and the 753-billion parameter GLM-5.2 model on a single workstation GPU.
Implications for Personal Computing
The researchers argue this work changes what personal machines can practically serve, turning open weights into deployable local software that makes frontier-scale intelligence accessible through hardware users already own. The system has been released publicly.