GLM-5.3-Flash served on single NVIDIA B200 with FreeToken

Developers tested running the 313-billion-parameter GLM-5.3-Flash mixture-of-experts model on a single B200 GPU with 191GB of available RAM by utilizing the FreeToken engine. Although the model checkpoint size of 198GB initially exceeded the GPU capacity, offloading experts to system memory allowed the system to successfully serve agentic sessions without crashing. The authors noted that this setup serves as a functional fallback rather than a high-performance serving path. Open article →