M5 Ultra vs. Mac mini M6
Our lab runs a Mac Studio M5 Ultra (2026) with 96 GB and a Mac mini M6 (2026). Connect your coding agent or chat in the browser and feel for yourself how fast large local AI models run on them – before you buy one.
Free: 2 chat questions and 2 minutes with your coding agent. No card needed.
The models
| Model | Chip | Generation | Reading | First token | Power while answering |
|---|---|---|---|---|---|
| Qwen3.8-Flash-Next 180B · 4-bit + MTP · 74.7 GB | Mac Studio M5 Ultra 96 GB | 80 tok/s | 3,270 tok/s | 0.12 s | 38 W |
| Mac mini M6 16 GB | does not fit | ||||
| Ornith-1.5 122B 122B · 4-bit · 65.5 GB | Mac Studio M5 Ultra 96 GB | 125 tok/s | 3,246 tok/s | 0.05 s | measurement coming |
| Mac mini M6 16 GB | does not fit | ||||
| Ornith-1.5 9B 9B · 4-bit · 4.8 GB | Mac Studio M5 Ultra 96 GB | 161 tok/s | 5,193 tok/s | 0.03 s | measurement coming |
| Mac mini M6 16 GB | 26 tok/s | 998 tok/s | 0.17 s | 7.6 W | |
A fair comparison needs the same model on both machines. Ornith-1.5 9B (4.8 GB) works because it fits on both. All figures are our own measurements.
Only on the Mac Studio M5 Ultra: Qwen3.8-Flash-Next (74.7 GB), Ornith-1.5 122B (65.5 GB). These models do not fit into the 16 GB of the Mac mini M6.
Last measured: November 10, 2026
Qwen3.8-Flash-Next
Qwen's 180B mixture-of-experts with 512 experts. Only a few billion parameters are active per token – that is what makes it fast.
Ornith-1.5 122B
Strong at code, reasoning and tool use, and faster per token than Flash-Next. Thinks before it answers.
Ornith-1.5 9B
A small model for comparison: it shows what fits into an ordinary Mac – and what you give up.
Does the model fit on my Mac?Rent a Mac Studio M5 Ultra by the hourOpenAI- and Anthropic-compatible API on a Mac Studio M5 UltraTry a Mac Studio M5 Ultra before you buy oneQwen3.8-Flash-NextOrnith-1.5 122BOrnith-1.5 9B
Two machines, both live
Test the hardware, not just the model: a Mac Studio with M5 Ultra for large models and a Mac mini with M6 as a compact comparison. Each machine has its own queue – you can use both in parallel.
Mac Studio M5 Ultra
Chip: M5 Ultra · CPU 30 cores · GPU 64 cores
Unified memory: 96 GB · For models: approx. 78 GB
Mac mini M6
Chip: M6 · CPU 12 cores · GPU 12 cores
Unified memory: 16 GB · For models: approx. 11.8 GB
FAQ
What is Try Silicon?
Two real Macs in our lab running open language models locally: a Mac Studio with M5 Ultra and 96 GB unified memory for large models, and a Mac mini with M6 and 16 GB for comparison. Test both live before you buy hardware or build your own setup.
What can I actually do here?
Chat in the browser, connect your coding agent to work on real projects, compare models and machines, and track tokens and speed in your dashboard. Typical uses: refactoring, writing tests, explaining code, drafting documentation, trying agent workflows.
Which coding agents work?
Anything that speaks the OpenAI API (/v1/chat/completions) or the Anthropic API (/v1/messages): Claude Code, opencode, the official SDKs and plain curl. Ready-made configs are on the agent page.
What is included for free?
2 questions in the chat and 2 minutes with your coding agent, starting with its first request – shared across both machines. When the time is up, your agent gets one last answer: what's done and what would still be open.
Why is there a queue?
To keep the numbers honest, each machine runs only one request at a time – every answer gets full speed. Each machine has its own queue, so you can e.g. let your agent work on the M5 Ultra and chat on the M6 at the same time. Paid requests go first; free requests move up after 90 seconds at the latest.
Is my code stored?
We do not store chats or prompts – only counts like tokens and speed. To make follow-up requests fast, the machine keeps a cache of recently processed text on its SSD, which is overwritten automatically. Please do not send secrets.
Ready to feel it?
No credit card required for the free test.