M5 Ultra vs M3 Ultra: Mac Studio local AI tests

NetworkChuck compares an Apple-loaned M5 Ultra Mac Studio with his M3 Ultra as machines to run models locally. His test units have 256 GB and 512 GB of unified memory respectively. The results are his own measurements, rather than independent verification of Apple's performance claims.

With the same 4-bit Qwen3.8 27B model and a 32,000-token summarization prompt, he reports time to first token of 21.66 seconds on the M5 and 82.85 seconds on the M3. Output reaches 43.3 versus 28.1 tokens per second. He connects the output speed difference to memory bandwidth and suggests neural accelerators explain the shorter prompt-processing wait. Further tests distinguish dense models from mixture-of-experts models, where parameter counts alone do not explain speed.

The workload tests extend beyond a local LLM. Whisper Large V3 transcribes 2 hours and 5 minutes of audio in a reported 89.4 seconds on the M5. He also tests footage descriptions, FLUX.2 klein 4B images and LTX-2.5 22B video through an MLX port.

Hermes Agent loads existing context with a local model, though he would reserve it for basic tasks rather than his main agent. Bionic from LM Studio connects a model to DaVinci Resolve through MCP and produces an edit, but the task takes about 13 minutes. The video contains no coding benchmark. The speaker says these workloads run offline; Micro Center sponsors the video.