DeepGrove · 2026-08-04 · major
Maple-Preview — 20B ternary MoE trained from scratch, 218 tok/s on a Mac mini
DeepGrove releases Maple-Preview, a 20B mixture-of-experts model trained natively in ternary weights (not post-quantized). MIT-licensed, 1B active per token, 24 layers, 256 experts with 8 active, 5.31 GB checkpoint, 218 tok/s decode on M4.
A 20B mixture-of-experts model trained in ternary from day one, small enough to run on a laptop.
Key specs
| Parameters | 20B total / 1B active |
|---|---|
| Experts | 256 (top-8) |
| Decode (m4) | 218 tok/s |
| Checkpoint | 5.31 GB |
Quick facts
| Maker | DeepGrove |
|---|---|
| License | MIT |
| Parameters | 20B total, 1B active |
| Architecture | 24-layer MoE, 256 experts, top-8 routing |
| Attention | 3:1 SWA-512 to global, 131K context |
| Quantization | Ternary (native, {-α, 0, +α}) |
| Decode speed | 218 tok/s on M4 Mac mini (Flash-head) |
What is it?
Maple-Preview from DeepGrove is a 20-billion-parameter mixture-of-experts language model where the weights are ternary — every parameter is one of three values, {-α, 0, +α}. Only about 1B parameters activate per token. The whole checkpoint is 5.31 GB, so it fits inside the memory of a laptop or an iPhone.
How does it work?
Unlike most low-precision models, Maple was trained directly in ternary form rather than converted from a full-precision checkpoint. It uses a 24-layer stack with 256 experts and top-8 routing, and a 3:1 mix of sliding-window (512 tokens) and global attention. A companion MLX fork, mlx-lm-deepgrove, streams and packs weights as 2-bit tensors so the 38 GB bf16 source never needs to fit in RAM.
Why does it matter?
Native ternary training is the design that finally makes 20B-class reasoning models practical on a phone or a Mac mini — 218 tok/s on M4, up to 395 tok/s on M5 Pro with Flash-head. It also gives developers an MIT-licensed base for on-device agents that need reasoning without a cloud round-trip.
Who is it for?
on-device developers, MLX users, reasoning-model researchers
Frequently asked questions
- What does 20B-A1B mean for Maple-Preview?
- Maple-Preview has 20 billion total parameters but only about 1 billion activate for any given token. That is the point of a mixture-of-experts design: a router picks a small subset of experts per token, so inference cost tracks the active count, not the full 20B.
- How is Maple-Preview different from a 2-bit quantized model?
- Maple-Preview was trained natively in ternary weights ({-α, 0, +α}) from day one, not converted afterwards. DeepGrove argues that post-hoc quantization loses accuracy because the model was optimized for higher precision, while training in the target precision keeps quality intact at the same footprint.
- What license is Maple-Preview released under?
- Maple-Preview is MIT-licensed. Both the model weights on Hugging Face and the mlx-lm-deepgrove inference fork on GitHub carry MIT, so commercial use, fine-tuning, and redistribution are all permitted with attribution.
- How fast can Maple-Preview run locally?
- Reported decode speeds are 218 tokens per second on an M4 Mac mini and 395 tokens per second on an M5 Pro with the Flash-head optimization enabled. The 5.31 GB checkpoint fits comfortably on Apple Silicon laptops and, per DeepGrove's post, on an iPhone as well.
Try it
pip install mlx-lm-deepgrove; huggingface-cli download deepgrove/maple-preview