Maple Preview
A 20B ternary MoE in 5.31GB, doing 218 tok/s on a Mac mini
What it is
Deepgrove's 20B-total / 1B-active mixture-of-experts built on ternary weights from the ground up: 24 layers, 256 experts with 8 active, a 3:1 SWA-512-to-global attention mix, and 131K context — a 5.31GB checkpoint that streams 218 tok/s on an M4 Mac mini through Deepgrove's mlx-lm fork, with a Triton/FlashAttention Transformers path for CUDA.
Why it's interesting
Bonsai proved you can crush a trained model down to ternary after the fact; Maple is the other bet — train ternary natively and let the MoE routing pay the quality bill. The card claims reasoning parity with its weight class on LCBv6, AIME 2026, HMMT 2026, and GPQA-D, and the whole thing is MIT with a LICENSE file in the repo.
Use cases
- Fast local reasoning on Apple Silicon without a discrete GPU
- Offline assistants where a 5GB footprint is the budget
- Research on native ternary training and low-bit MoE routing
Who it's for
Local-AI tinkerers, Apple Silicon users, quantization researchers
Setup
Easy. ~6GB free memory; Apple Silicon via the mlx-lm-deepgrove fork, or CUDA with Triton and FlashAttention. Community GGUFs are appearing
Limitations & cautions
It's a preview with minimal agentic post-training and only small-scale RL, the benchmark claims are the card's own with no independent numbers yet, and Deepgrove is a young outfit — treat the quality story as promising, not proven.
Editorial takeaway
Bonsai compresses a big model into a phone. Maple asks the sharper question: what if it was never big to begin with?