Found first: a primary source the press has not covered yet.
A joint team from MIT and UC Berkeley has published FreeToken, a serving system for mixture-of-experts models that continuously remaps computation and storage to match available hardware bandwidth. The paper, submitted August 17, 2026, reports 753B-parameter GLM-5.2 running on a single workstation GPU.
What the source says
FreeToken co-designs the full serving stack: model layout and loading, expert residency, CPU-GPU execution, agentic state reuse, and runtime memory management. Rather than fixed offloading schedules, it dynamically reassigns which experts reside in GPU memory based on current bandwidth conditions. Reported hardware configurations include a laptop GPU with 8 GB of VRAM running 35B-parameter models, gaming desktop hardware running 284B-parameter models, and a single workstation GPU running 753B GLM-5.2. The system supports more than 20 MoE models and was tested on real coding and tool-using agent workloads. Authors are Shuo Yang, Xiaoze Fan, Melissa Pan, Haocheng Xi, Zhe Wang, Shanlin Sun, Kurt Keutzer, and Song Han (MIT), and Matei Zaharia, Chenfeng Xu, and Ion Stoica (UC Berkeley).
Why it matters
Running models at the 700B-plus scale has required server-grade GPU clusters. A result at that scale on a single workstation GPU, using bandwidth-adaptive scheduling rather than additional hardware, changes what solo researchers and small teams can run locally. The author group spans two of the most active inference systems labs: Song Han's hardware-aware ML group at MIT and the Berkeley Sky Computing Lab, which produced vLLM. Testing against coding and tool-using agent workloads rather than synthetic throughput benchmarks makes the serving figures directly interpretable for practitioners building agentic pipelines. Code and models are available at flashml.ai.