deepgrove/maple-preview
Summary
DeepGrove releases Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM with SOTA reasoning for its weight class, capable of 200+ tokens/sec on a Mac mini M4 and competitive with larger models.
View Cached Full Text
Cached at: 08/05/26, 07:39 PM
deepgrove/maple-preview · Hugging Face
Source: https://huggingface.co/deepgrove/maple-preview DeepGrove · 2026
Today we introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM. Maple-Preview has SOTA reasoning for its weight class and is even competitive with larger models. It solves IMO-level problems and runs at 200+ tokens/sec on a Mac mini M4, 5–16× faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss.
- 20B-A1B Model
- 218 tok/s M4 Mac mini
- 5.31 GB Checkpoint
- 131,072 Token context
The included Transformers implementation depends on Triton and FlashAttention and is intended for a compatible CUDA environment. The reported Apple Silicon result uses a separate on-device runtime.
https://huggingface.co/deepgrove/maple-preview#architectureArchitecture
Maple-Preview is a 20B-A1B reasoning model designed from the start for efficient on-device inference. It utilizes a 24-layer, 256-expert (8 active) configuration with 3:1 SWA-512:GA attention.
https://huggingface.co/deepgrove/maple-preview#evaluationEvaluation
On benchmarks, Maple-Preview sets a new point on the Pareto frontier for both memory-to-performance and speed-to-performance, demonstrating its strong reasoning capabilities. However, we note that this preview is focused primarily on raw reasoning and, as such, may underperform on agentic benchmarks. We intend to continue improving general performance through extended training before Maple’s full release.
Capability comparison using the dense output head across LCBv6, AIME 2026, HMMT 2026, and GPQA-D.
https://huggingface.co/deepgrove/maple-preview#limitationsLimitations
This preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning.
https://huggingface.co/deepgrove/maple-preview#licenseLicense
Maple-Preview is released under theMIT License.
Similar Articles
Show HN: Maple-Preview – ternary 20B MoE running at 120 tok/s on a iPhone
Maple-Preview is a ternary 20B MoE model that runs at 120 tokens per second on an iPhone, showcasing efficient on-device inference.
I've added Maple-Preview to Mference, got 40 tps generation with 500MB of used RAM on Air M4
The author adds Maple-Preview support to Mference, a tool that streams MoE experts from disk to run large models on low-RAM devices, achieving 40 tps with 500MB RAM on an Air M4.
Microsoft's new MAI models
Microsoft announced two new LLMs: MAI-Thinking-1 (35B reasoning model) and MAI-Code-1-Flash (5B code model), both trained on enterprise-grade, clean data without third-party distillation, with MAI-Thinking-1 claimed to be preferred over Sonnet 4.6 in blind evaluations.
VibeThinker: 3B param model that beats Opus 4.5 on reasoning with novel SFT+GRPO
This technical report introduces VibeThinker-3B, a 3B parameter dense model that achieves frontier-level reasoning performance on benchmarks like AIME26 and LiveCodeBench, matching or exceeding much larger models such as DeepSeek V3.2 and GLM-5 through a combination of curriculum-based SFT, multi-domain RL, and offline self-distillation.
@jinyuhou0: On popular benchmarks, our 30B model matches systems 20-30x its size (gpt-5.4-xhigh, DeepSeek-V3.2, Kimi-K2.5), while u…
A new 30B model matches systems 20-30x its size on popular benchmarks while using up to 95% fewer reasoning tokens than comparable agentic LLMs, achieved through a learned configurator that decides when and how to reason. Model and code are openly available.