Run Qwen 3.8 Flash Next on your own Mac, with a lean coding agent
On my 128 GB M3 Max, Qwen 3.8 Flash Next built a working app from one prompt in about 13 minutes. The biggest speedup was the agent setup, not the model.
Why it matters
I ran the 94 GB version (UD-IQ4_XS) with llama.cpp on a 128 GB M3 Max MacBook Pro. It wrote about 18 tokens per second.
With my skills and extensions on, the coding agent sent about 20,000 tokens before the model wrote a word. A local model reads a long prompt slowly.
In Pi, the flags -ne -ns -nc turn off extensions, skills and AGENTS.md files. In my test the first prompt went from 19,567 tokens to 1,684.
Do this
- Check your RAM: the 128 GB row of the guide uses the 93.7 GB file. Under 96 GB, this model does not fit.
- Mac only: let the GPU use more memory with sudo sysctl iogpu.wired_limit_mb (the exact line is in the guide, Step 3).
- Start the server with the llama-server command below. The first start downloads the model.
- In a second terminal, add the local provider to Pi (python3 add-to-pi.py in the guide folder), then start Pi with the lean flags below.
- Check the bottom right of Pi shows qwen3.8-flash-next and no $ cost, so you are on the local model.
The two commands
llama-server -hf unsloth/Qwen3.8-Flash-Next-GGUF:UD-IQ4_XS --alias qwen3.8-flash-next -c 65536 -ngl 99 -fa on --jinja --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 --host 127.0.0.1 --port 8080 pi -ne -ns -nc --provider llama-cpp-flash-next --model qwen3.8-flash-next
One thing to keep in mind: The lean flags also turn off your own skills and project rules, so the agent knows less about your repo. Use them for a fresh folder or a small task.
Numbers are from my own test on camera (YouTube: Qwen 3.8 Flash Next Tested on a 128GB Mac). File sizes are from the Hugging Face file list, read 2026-11-09. Source: Full step-by-step guide on GitHub.
Free live lesson · Nov 24Stop Measuring AI by Tokens. Measure What It EarnsOne hour, live, with Saurav. Bring your own bill. Free.Save my free seat → Ask me anything about AI and cloudA paid call on your own setup: your agents, your bill, your plan.Book a call →