What changed
Apple's updated Mac Studio with the M5 Ultra SoC has demonstrated superior agentic AI performance compared to the DGX Spark system, particularly in tasks involving prompt processing and decode speed. The M5 Ultra's 80-core GPU and 1.2 TB/s memory bandwidth significantly outperform the DGX Spark's 48 Blackwell SMs and 273 GB/s memory bandwidth.
What this means in practice
The Mac Studio's ability to run large models like DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash efficiently highlights its potential for local AI workloads. While the M5 Ultra can fit quantized versions of these models in its 256GB of RAM, the DGX Spark system requires clustering multiple units to achieve similar results. The Mac Studio finishes agentic tasks in under half the time of the DGX Spark, with a 16% lower energy consumption. However, the Mac Studio's performance with DeepSeek V4.1 Flash is limited by the lack of compatible inference engines and memory constraints, making the DGX Spark cluster a better option for this model.
Limitations and availability
The Mac Studio's agentic AI performance is highly dependent on the inference engine used. While llama.cpp works well for some models, it does not support the DSpark speculative decoding model required for optimal performance with DeepSeek V4.1 Flash. The DGX Spark cluster, using TensorFold and a community recipe, delivers faster results for this model. The Mac Studio's 256GB of RAM is sufficient for running quantized versions of large models, but it cannot handle both the model weights and the decoding model simultaneously. The Mac Studio is currently available for purchase, though with a long lead time of 13 to 15 weeks. The DGX Spark system, while powerful, is significantly more expensive and less accessible due to high costs and limited availability.
