setup 12reel 422026-08-29comment MAPLE
Run a 20B AI model in 7GB RAM
Maple-Preview, a 20B-A1B ternary MoE, runs at 218 tokens/sec with 6.69GB peak RAM on an M4 Mac mini.
This is the setup and the honest catch for Maple-Preview, the 20B-A1B ternary MoE from DeepGrove.
Everything below is from the Hugging Face model card and the mlx-lm-deepgrove repo, checked on 29 August 2026.
What it actually does: Maple-Preview is a reasoning model packed into a 5.31GB checkpoint. DeepGrove reports it running at 218 tokens/sec decode on an M4 Mac mini with 6.69GB peak RAM using --flash-head.
The numbers
| Receipt | Value |
|---|---|
| Model | Maple-Preview |
| Architecture | 20B-A1B ternary MoE |
| Layers and experts | 24 layers, 256 experts, top-8 |
| Checkpoint | 5.31GB |
| Source model | 38GB bf16 |
| Context | 131,072 tokens |
| M4 peak RAM | 6.69GB with --flash-head |
| M4 decode speed | 218 tokens/sec with --flash-head |
| License | MIT |
Rounded for the reel: 7GB RAM.
What to install
The fast local path is the MLX fork, so this is for Apple Silicon.
git clone [email protected]:deepgrove-ai/mlx-lm-deepgrove.git
cd mlx-lm-deepgrove
./setup.sh
source .venv/bin/activate
hf download deepgrove/maple-2bit-mlx --local-dir maple-2bit-mlx
Run it:
python -m mlx_lm chat --model ./maple-2bit-mlx --trust-remote-code --max-tokens -1 \
--temp 1.0 --top-p 0.95 --flash-head
The catch
The viral version says iPhone.
I would not use that as the setup claim yet. The repo benchmark lists Macs, not iPhones, and says the runtime requires Apple Silicon. Until there is a reproducible iPhone run, the honest claim is:
Run a 20B model in about 7GB RAM on Apple Silicon.
That is still absurd enough.
What it is good at
Maple-Preview is a reasoning preview, not a polished assistant. The model card says it is focused on raw reasoning and may underperform on agentic benchmarks.
So do not treat it like a full ChatGPT replacement. Treat it like a tiny, fast reasoning model you can actually run locally.
Sources
- Hugging Face model card: huggingface.co/deepgrove/Maple-Preview
- MLX runtime and benchmark: github.com/deepgrove-ai/mlx-lm-deepgrove
If the repo changes after 29 August 2026, trust the repo over this page.
Which Mac are you running it on?
Ask me in The Rabbit Hole and I will tell you whether to expect anything near that 218 tokens a second.
It is the room where I keep the parts that don't fit on a page like this: the full run of numbers, what broke the first time, and the ones that looked good and then didn't work at all.
Two catches, up front, because that's the rule here. It's private, so you have to be inside to read anything. And it's new, so you'd be early rather than joining something busy.
if you run it, tell me
the rabbit hole
It's free to join, and the course library is free too. Ask there and I answer in the open, so the next person with the same question finds it. There's an optional paid Premium tier for weekly office hours.
join the rabbit holePost in there what you ran it on and what you got, or reply to the DM. I read every one.