User Benchmarks GLM-5.3-Flash on Mac Studio M5 Ultra with DwarfStar Engine
A user shared performance observations from running the GLM-5.3-Flash model using the DwarfStar inference engine on an 80-core Apple M5 Ultra Mac Studio with 256GB of unified memory. The report highlights that while the system easily holds large models, GPU compute performance becomes a clear bottleneck during heavy agentic inference tasks. Highlighting hardware bottlenecks in high-RAM Apple Silicon setups helps developers and AI enthusiasts make informed decisions when purchasing systems for local LLM inference. It underscores the distinction between unified memory capacity needed to fit large models and raw GPU compute power required for high token generation speeds. The user noted that although 256GB of unified memory successfully runs massive models, scaling to a 512GB configuration may offer diminishing returns because GPU processing speed remains the limiting factor. DwarfStar is a specialized, lightweight inference engine designed to optimize model execution via Metal on Apple hardware and CUDA on GPUs.
## BACKGROUND
Apple Silicon Mac Studios utilize unified memory architecture (UMA), allowing the GPU to directly access up to hundreds of gigabytes of RAM shared with the CPU, making them popular choices for running large language models locally. DwarfStar is an open-source inference engine designed for native optimization on Apple Metal and CUDA, targeting models like DeepSeek V4 and GLM-5.3 Flash.