Maple Preview: DeepGrove's Ternary Local LLM

Maple-Preview is a 20B-A1B ternary-weight reasoning LLM

Maple-Preview is an open-source reasoning LLM from DeepGrove, billed as 20B-A1B and built on ternary weights. DeepGrove's model card says it has the best reasoning in its weight class and is competitive with larger models. Both are the developer's claims, and every figure about the model comes from DeepGrove's own model card on Hugging Face.

What 20B-A1B and ternary weights mean

The card never defines 20B-A1B; it gives the layout instead: 24 layers and 256 experts, 8 of them active. The attention is listed as 3:1 SWA-512:GA. Experts are the parallel sub-networks a mixture-of-experts layer chooses among; the Hugging Face page tags the model as mixture-of-experts.

The card does not say whether "8 active" counts experts per token or per layer. "Ternary-weight" is likewise the card's own term, with no definition and no list of the values a weight can take. The repository page carries a ternary tag, which confirms the label but adds nothing more.

The 5.31 GB checkpoint and 131,072-token context

The card's stat block lists a 5.31 GB checkpoint and a 131,072-token context. A checkpoint is the saved file of trained weights, and context is how many tokens the model can take in at once.

The card does not say which file the 5.31 GB describes. The Hugging Face page lists the stored tensor type as BF16, so check the repository's file listing for actual sizes before you plan disk space around that figure.

218 tokens per second on a Mac mini M4, with a caveat

DeepGrove cites 200+ tokens per second on a Mac mini M4, but the card ties that result to a separate on-device runtime. Don't expect that speed from the code that ships with the model. The opening paragraph also says the model solves IMO-level problems, and it sets the speed at 5–16x that of Gemma 4, Qwen3.5 and gpt-oss.

The stat block gives 218 tok/s for the same machine. Both figures are DeepGrove's own.

The included code targets CUDA, not Apple Silicon

The shipped Transformers code is meant for CUDA, and the Apple Silicon number comes from a different runtime. CUDA is NVIDIA's GPU runtime, and Apple Silicon Macs do not have NVIDIA GPUs. The card names Triton and FlashAttention as dependencies of the included implementation.

The speed figure and the shipped code therefore describe two different setups. The card does not say whether that runtime is available to the public, and it gives no memory or hardware requirement for a Mac either.

For local LLM coding, DeepGrove flags weaker agentic results

The card makes no coding claim; it warns that a preview built for raw reasoning may fall short on agent benchmarks. In this context, an agent is a model that calls tools and acts on the results over several steps. Beyond that warning, the card's prose says nothing about coding quality.

Reasoning results: the Pareto frontier claim

DeepGrove claims a new point on the Pareto frontier, for memory against performance and for speed against performance. A Pareto frontier marks the best trade-offs on offer: a point on it is one no other model beats on both axes at once. That is DeepGrove's reading of its own benchmarks, and the card cites no independent rerun.

Minimal agentic post-training ahead of the full release

The Limitations section gives the cause of the agentic gap: little post-training for agent tasks and only a small round of general reinforcement learning. Post-training is the extra training after the main run, and reinforcement learning shapes a model by rewarding good outputs.

DeepGrove says more training will come before Maple's full release. The card frames the weaker agentic results as a limit of the preview stage. If you want a small local model for agent work, Liquid AI LFM2.5-2.6B: Fast Agents, Not for Coding covers one whose maker pitches it at agents and advises against using it for coding.

MIT license and download: what to check before using it

The card says Maple-Preview is released under the MIT License, and the model is published on Hugging Face as deepgrove/maple-preview. The card names the license but does not spell out what it permits. Read the repository's LICENSE file for the terms before any commercial use.

Two things the card leaves open are worth checking before you download: how the repository's file sizes compare with the 5.31 GB figure, and whether you can obtain the runtime behind the Apple Silicon result. The model is also labeled a preview, and DeepGrove plans more training before the full release.