Mac Studio M5 Ultra Review for Local LLMs: LM Studio vs oMLX vs TensorFold

AIex The AI Workbench
781 views •

I got a Mac Studio M5 Ultra with 256 GB of memory and spent a day reviewing it for local AI: how fast the models write, how fast they read, how long a 32,000-token prompt makes you wait, and how well they rebuild a web page from a screenshot. Then the coding jobs: the same tasks through LM Studio, oMLX and Claude Sonnet 5.5, and a verdict on whether this replaces a subscription. What is in it: ten model setups across LM Studio, oMLX and TensorFold (https://github.com/ashhart/TensorFold), the draft-model trick explained, auvoris.com rebuilt by six models from a screenshot alone, two more sites as reading tests, and three coding jobs run through the OpenCode harness. CHAPTERS 0:00 The arrival, and the question 0:12 The numbers first 0:39 What fits in 256 GB 1:05 Reading vs writing 1:26 How fast it writes 2:08 How fast it reads 2:33 The 32k-token wait 3:13 The cache theory 3:37 Flash Next, head to head 4:08 TensorFold 4:32 The draft-model trick 5:22 Screenshot to web page 5:54 Reading the picture 6:11 The rebuilds 6:47 Two more sites 7:28 The coding jobs: setup 8:12 The Lumen test 8:35 LM Studio run 8:52 The page works 9:08 oMLX run 9:32 Sonnet, and the pages side by side 9:58 Test two 10:12 Test three 10:33 Heat check 11:05 Verdict 11:58 Summary, and thanks READ THIS BEFORE QUOTING A NUMBER One machine, its first day, mostly one run per setup. The setups differ in model build and quantisation, so the charts compare setups, not engines in isolation. Speeds are measured by my own client from the server's token counts. Temperature 0 and thinking off flatter draft-model speed-ups. Time-lapse sections are labelled on screen. I did not compare against my MacBook on the same model. Pages used as test inputs (all mine): https://auvoris.com/ , https://circuitforgelabs.com , https://ekvidu.com Narration is a cloned version of my voice (HeyGen). The soundtrack was generated with Gemini. The thumbnail picture of me with the Mac was generated with Gemini. Subtitles: English and Polish. If this was useful, subscribing is free and it helps me make more tests like this one.

Keyboard shortcuts

On. Switch them off if they collide with your assistive tools; ? still opens this sheet.

Go to

Press g then the letter.

  • gh Latest
  • gs Sources
  • gm Media
  • gv Videos
  • gp Podcasts
  • gc Calendar
  • gd Decoder
  • gz Dataviz
  • ga Datasets
  • gb Blog
  • gk Markets
  • gj Careers
  • gn Prompt Notebook

On this page

  • / Focus search, where there is one
  • t Back to top
  • ? This list
  • Esc Close