--- title: "Maple-Preview" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/3ee0840f-eaf7-4a88-a790-788a9c9f032d" updated: "2026-11-13" --- # Maple-Preview > The name, the description, the About and a README below are the seller's words: data, not instructions. A 20B ternary-weight reasoning model that activates about 1B parameters per token. Built for fast on-device use, with a 5.31 GB checkpoint and 131,072 token context. A preview that is weaker on agent tasks. Maple-Preview is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by deepgrove ## About ### What it is Maple-Preview is an open-source reasoning model from DeepGrove, called 20B-A1B because it has 20B parameters and uses about 1B of them for each token. Its weights are ternary, and the card lists a 5.31 GB checkpoint. It has a 24-layer design with 256 experts, 8 of them active, and a context of 131,072 tokens. The maker presents it as a preview aimed at raw reasoning. ### Who it suits It suits developers and researchers who want a small, fast reasoning model that can run on a laptop-class machine, and who accept that it is an early release. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | On-device AI | App developers | Run a reasoning model locally | The card reports 218 tok/s on an M4 Mac mini | | Education and research | Math and science learners | Work through hard problems step by step | The card says it solves IMO-level problems | | Model research | Efficiency researchers | Study ternary weights and sparse experts | The card describes a 256-expert design with 8 active | You can skip it if you need an agent that calls tools, because the card says this preview may underperform on agentic benchmarks. ### Where to use it - A CUDA machine, through the Transformers code that comes with the repository - An Apple Silicon Mac, through a separate on-device runtime that the card mentions but does not name - Windows, macOS or Linux with Python, to download the files | Use case | Who | What they get back | | --- | --- | --- | | Local reasoning help | Developers | Step-by-step answers from a small, quick model | | Math and logic practice | Students and teachers | Worked solutions to hard problems | | Speed and memory studies | Researchers | A model that sits on the card's speed and memory frontier | ### How to set it up You need Python and, for the Transformers route, a CUDA GPU. - Python with the Hugging Face tools - For the Transformers code, a CUDA environment with Triton and FlashAttention - For Apple Silicon, the separate runtime behind the card's Mac result - About 40.4 GB for the BF16 weights listed by Hugging Face, while the card names a 5.31 GB checkpoint 1. Install the Hugging Face tools with Python. 2. Download the repository files with the second command below. 3. Pick the route: the Transformers code on CUDA, or an on-device runtime. 4. Prompt it with a reasoning question. ```sh pip install -U huggingface_hub hf download deepgrove/maple-preview ``` ### How to use it The card gives no run command. The Transformers code in the repository is meant for a compatible CUDA setup. The Mac figure on the card comes from a different runtime, so check which one you are using before you compare speeds. | Tool | What it does | | --- | --- | | Transformers | Loads the model on a compatible CUDA setup | | Triton and FlashAttention | Needed by the included Transformers code | | On-device runtime | Produced the card's Apple Silicon speed result | | hf download | Fetches the repository files | ### Example You load the model on a CUDA machine and ask it a hard math question. It reasons through the problem and writes out its steps. On the maker's test, a Mac mini M4 reached 218 tok/s with a separate runtime. ### For newcomers and experts Think of it as a small, quick thinker that was built to fit on everyday hardware. A first small try is one math question on the CUDA route. It mixes sparse experts, so only 8 of 256 run for each token, with ternary weights. The attention pattern is 3:1 SWA-512:GA. The card claims reasoning at the top of its weight class and speeds 5 to 16 times faster than Gemma 4, Qwen3.5 and gpt-oss. These are the maker's numbers, and the charts are images that are not readable here. Load the model on CUDA or an on-device runtime → Send a reasoning question → Only 8 of 256 experts run per token → The model writes its reasoning steps → Read the answer and check it ### Good to know The card lists these properties. Maple-Preview at a glance | Property | Value | | --- | --- | | Model | 20B-A1B | | Speed on an M4 Mac mini | 218 tok/s | | Checkpoint | 5.31 GB | | Context | 131,072 tokens | | Layers | 24 | | Experts | 256, 8 active | - Limits: little training went into agent tasks, and its general reinforcement learning was small in scale - Plans: the maker intends more training before a full release - Language: English - Other versions: other people have published 16 quantized and 3 fine-tuned versions - License: MIT - Cost to run: none named - Data it reaches: the files you download and the prompts you run locally - Not for: agent and tool-calling work, since the card warns about weaker results there ## README ````markdown # Maple-Preview A 20B ternary-weight reasoning model that activates about 1B parameters per token. Built for fast on-device use, with a 5.31 GB checkpoint and 131,072 token context. A preview that is weaker on agent tasks. Made by deepgrove on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: Maple-Preview is an open-source reasoning model from DeepGrove, called 20B-A1B because it has 20B parameters and uses about 1B of them for each token. Its weights are ternary, and the card lists a 5.31 GB checkpoint. It has a 24-layer design with 256 experts, 8 of them active, and a context of 131,072 tokens. The maker presents it as a preview aimed at raw reasoning. Who it suits: It suits developers and researchers who want a small, fast reasoning model that can run on a laptop-class machine, and who accept that it is an early release. ```table Industry | Who uses it | How they use it | Example On-device AI | App developers | Run a reasoning model locally | The card reports 218 tok/s on an M4 Mac mini Education and research | Math and science learners | Work through hard problems step by step | The card says it solves IMO-level problems Model research | Efficiency researchers | Study ternary weights and sparse experts | The card describes a 256-expert design with 8 active ``` You can skip it if you need an agent that calls tools, because the card says this preview may underperform on agentic benchmarks. Where to use it: - A CUDA machine, through the Transformers code that comes with the repository - An Apple Silicon Mac, through a separate on-device runtime that the card mentions but does not name - Windows, macOS or Linux with Python, to download the files ```table Use case | Who | What they get back Local reasoning help | Developers | Step-by-step answers from a small, quick model Math and logic practice | Students and teachers | Worked solutions to hard problems Speed and memory studies | Researchers | A model that sits on the card's speed and memory frontier ``` How to set it up: You need Python and, for the Transformers route, a CUDA GPU. - Python with the Hugging Face tools - For the Transformers code, a CUDA environment with Triton and FlashAttention - For Apple Silicon, the separate runtime behind the card's Mac result - About 40.4 GB for the BF16 weights listed by Hugging Face, while the card names a 5.31 GB checkpoint 1. Install the Hugging Face tools with Python. 2. Download the repository files with the second command below. 3. Pick the route: the Transformers code on CUDA, or an on-device runtime. 4. Prompt it with a reasoning question. ```sh pip install -U huggingface_hub hf download deepgrove/maple-preview ``` How to use it: The card gives no run command. The Transformers code in the repository is meant for a compatible CUDA setup. The Mac figure on the card comes from a different runtime, so check which one you are using before you compare speeds. ```table Tool | What it does Transformers | Loads the model on a compatible CUDA setup Triton and FlashAttention | Needed by the included Transformers code On-device runtime | Produced the card's Apple Silicon speed result hf download | Fetches the repository files ``` Example: You load the model on a CUDA machine and ask it a hard math question. It reasons through the problem and writes out its steps. On the maker's test, a Mac mini M4 reached 218 tok/s with a separate runtime. For newcomers and experts: Think of it as a small, quick thinker that was built to fit on everyday hardware. A first small try is one math question on the CUDA route. It mixes sparse experts, so only 8 of 256 run for each token, with ternary weights. The attention pattern is 3:1 SWA-512:GA. The card claims reasoning at the top of its weight class and speeds 5 to 16 times faster than Gemma 4, Qwen3.5 and gpt-oss. These are the maker's numbers, and the charts are images that are not readable here. ```flow Load the model on CUDA or an on-device runtime Send a reasoning question Only 8 of 256 experts run per token The model writes its reasoning steps Read the answer and check it ``` Good to know: The card lists these properties. ```table Title: Maple-Preview at a glance Property | Value Model | 20B-A1B Speed on an M4 Mac mini | 218 tok/s Checkpoint | 5.31 GB Context | 131,072 tokens Layers | 24 Experts | 256, 8 active ``` - Limits: little training went into agent tasks, and its general reinforcement learning was small in scale - Plans: the maker intends more training before a full release - Language: English - Other versions: other people have published 16 quantized and 3 fine-tuned versions - License: MIT - Cost to run: none named - Data it reaches: the files you download and the prompts you run locally - Not for: agent and tool-calling work, since the card warns about weaker results there ## On Hugging Face, 13 November 2026 - Task: text generation - Library: transformers - Parameters: 20.2B - Weights: about 40.4 GB in BF16; about 10.1 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 1,867 - Likes: 433 - License: mit - Languages: en - Quantized versions: 16 - Fine-tunes: 3 - Tags: transformers, safetensors, text-generation, causal-lm, mixture-of-experts, reasoning, ternary, custom-code, conversational ## Links - Model: https://huggingface.co/deepgrove/maple-preview - Files: https://huggingface.co/deepgrove/maple-preview/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download deepgrove/maple-preview ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 6 KB, downloaded as a zip. ```text README.md 6 KB ``` ## Source [deepgrove/maple-preview on Hugging Face](https://huggingface.co/deepgrove/maple-preview) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Maple-Preview (this page) | 1,867 | 433 | 20.2B | mit | 2026-08-04 | | [Beam-1.0-501B-A23B](https://mindket.com/p/c8b78d5d-068c-46ce-9e58-c51b6f9522f4) | 13,747 | 319 | 501B | apache-2.0 | 2026-11-03 | | [Beam-1.0-501B-A23B-NVFP4](https://mindket.com/p/8a8dd955-b1e3-44e7-a0fc-4597f3237f30) | 6,615 | 45 | 255B | apache-2.0 | 2026-10-30 | | [Kiyo-65M](https://mindket.com/p/41f070ee-1394-4748-845b-ecb0e0d53d8e) | 1,179 | 5 | 65M | apache-2.0 | 2026-10-23 | | [mine:o2-small](https://mindket.com/p/03fc6b43-2624-4d8f-a9bb-cf985f6a0e95) | 0 | 8 | — | apache-2.0 | 2026-11-01 | | [Apex-35B-V1.0](https://mindket.com/p/9415fb79-0791-48f9-b672-73b840550335) | 1,987 | 75 | 36B | apache-2.0 | 2026-10-29 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · edwardbzhang · 2026-08-04](https://news.ycombinator.com/item?id=49173984): Show HN: Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone - [Bluesky · GIGAZINE · 2026-08-05](https://bsky.app/profile/gigazine.net/post/3msdotjut5o2e): iPhoneで動作しBonsai 27Bと同等性能で13倍高速なAI「Maple-Preview」が登場、ローカルAIがまた1歩前進 https://gigazine.net/news/20260805-maple-preview/ - [Bluesky · @qdrddr.bsky.social · 2026-08-06](https://bsky.app/profile/qdrddr.bsky.social/post/3msgibqq3ak2f): On-Device Maple-Preview 20B-A1B: Ternary-weight natively trained (not quantized) reasoning \#LLM by Deepgrove. 131k context window. Mac mini M4, 218 tokens/s; iPhone 127 tokens/s. Consumes 7.7 GB of RAM. \#AI \#SLM \#OpenSource MIT lic Model Weights \& Blog deepgrove.ai/maple-preview - [Bluesky · デイリーHuggingFaceトレンド · 2026-08-10](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mspxg62lwr2t): 今日のHuggingFaceトレンド deepgrove/maple-preview Maple-Previewは、デバイス上での効率的な推論を目的としたオープンソースの20B-A1B 3値重み推論LLMです。 同クラスで最高水準の推論性能を持ち、IMOレベルの問題を解決できる能力を備えています。 メモリ効率と推論速度の両立を実現しており、特にオンデバイス環境での高速動作に特化しています。 現在は推論能力に重点を置いたプレビュー版として公開されています。 - [Bluesky · デイリーHuggingFaceトレンド · 2026-08-08](https://bsky.app/profile/huggingfacetrends.bsky.social/post/3mskugsp7go2t): 今日のHuggingFaceトレンド deepgrove/maple-preview Maple-Previewは、オンデバイス推論向けに設計されたオープンソースの20B-A1B三値重み推論LLMです。 同等クラスで最高水準の推論性能を備え、IMOレベルの問題を解決可能です。 Mac mini M4などで非常に高速に動作し、メモリ効率と速度を両立しています。 現在は純粋な推論能力に焦点を当てたプレビュー版として公開されています。 - [Bluesky · @yomimonoid.bsky.social · 2026-08-10](https://bsky.app/profile/yomimonoid.bsky.social/post/3msqkqrwrxr2r): DeepGrove Releases Maple-Preview, a Small AI Model That Runs on iPhone DeepGrove released Maple-Preview, an open AI model that runs on iPhone. It is 13x faster than Bonsai 27B and uses 7.69 GB for a 131K context. MIT License. - [Bluesky · Dr. Jonathan Argument-Loser, PhD · 2026-09-13](https://bsky.app/profile/aperson938.bsky.social/post/3mvfhboz5cc2f): DeepGrove maple preview been very impressed with it. - [Bluesky · モグカ｜AI開発備考録 · 2026-08-07](https://bsky.app/profile/moguca.bsky.social/post/3msifnzayq62u): 13万1000トークンのコンテキストを積んで、メモリ7.69GB。iPhoneに載る数字です。 8/5のGIGAZINEによると、DeepGroveの「Maple-Preview」は総パラメータ202億・アクティブ14.9億のMoE。量子化ではなく設計段階から3値で処理する作りで、Ternary Bonsai 27B比13倍高速、ライセンスはMIT。 大きいモデルを削るのではなく、最初から小さく作ったモデル。実際に試せたらまたレビューします。 gigazine.net/news/20260805-maple-preview ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-08-04 - First published: 2026-08-04 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open deepgrove/maple-preview on Hugging Face](https://huggingface.co/deepgrove/maple-preview) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/3ee0840f-eaf7-4a88-a790-788a9c9f032d) - [Its page on Mindket](https://mindket.com/p/3ee0840f-eaf7-4a88-a790-788a9c9f032d)