Maple-Preview 20B: DeepGrove's Ternary Reasoning Model Explained

What You'll Learn

• What Maple-Preview is and why ternary weights matter for on-device AI
• Verified benchmark scores across LiveCodeBench v6, AIME 2026, HMMT 2026 and GPQA-Diamond
• How to run it locally on CUDA, Apple Silicon or via the official GGUF build
• Its real limitations as a preview release and where DeepGrove's roadmap goes next

Maple-Preview is a brand-new open-source reasoning model that changes what small models can do on everyday hardware. DeepGrove, a San Francisco startup backed by Y Combinator, released it on August 4, 2026, and the weights are live on Hugging Face under the MIT license. The headline numbers are hard to ignore: a 20B-A1B mixture-of-experts design that solves IMO-level math, a 5.31 GB checkpoint, a 131,072-token context window, and 218 tokens per second on a Mac mini M4. According to the official model card, that makes it five to sixteen times faster than efficient models such as Gemma 4, Qwen3.5 and gpt-oss. Maple-Preview is built on ternary weights - parameters limited to -1, 0 and +1 - trained from scratch instead of compressed after the fact. This guide covers what it is, how it performs, how to run it locally, and where the preview still falls short.

What Is DeepGrove's Maple-Preview?

Maple-Preview is a 20B-A1B reasoning model designed from the start for efficient on-device inference. The name encodes the architecture: roughly 20 billion total parameters with about 1.5 billion active per token, because only 8 of 256 experts fire at a time. That sparse mixture-of-experts layout is what lets a "20B" model behave like a small one at runtime, both in memory and in compute.

DeepGrove is a three-person San Francisco startup whose Y Combinator profile describes the mission as "frontier intelligence on any device." Harj Taggar, a managing partner at Y Combinator, is the company's primary partner and publicly championed the release. The team trained the model in ternary format from the very beginning rather than converting a full-precision model later, which is why the checkpoint stays so small. For the on-device segment of the market - a space we mapped in detail in our Edge AI 2026 guide - that combination of size, speed and license terms is unusual.

Why Ternary Weights Make Maple-Preview Different

Most open models ship weights in 16-bit or 8-bit precision. Maple-Preview restricts every weight to one of three values: -1, 0 or +1. That ternary format takes roughly 2 bits per weight instead of 16 or 32, and research on ternary LLMs shows it can cut model size by around 16 times compared with 32-bit storage while keeping accuracy close to a full-precision model of the same size.

The bigger trick is arithmetic. When weights are limited to -1, 0 and +1, most float multiplications collapse into simple additions and subtractions. Inference becomes memory-bound instead of compute-bound, so the bottleneck is how fast the device can read the weights, not how fast it can multiply. DeepGrove says this design drops the memory read per token to about 24 MB and lets the whole model footprint fit inside 16 GB of RAM. It sits alongside other open-weight efforts like Cohere's Aya Expanse family, but it is one of the first mainstream releases trained natively in ternary format rather than quantized afterward.

Maple-Preview Benchmarks and Specs

DeepGrove publishes capability numbers for a dense output head across four tough evaluations, and independent trackers such as Benchgen report the same scores. The average across the four benchmarks is 78.7%, which is strong for a model with roughly 1.5 billion active parameters. Maple-Preview also solved all seven IMO 2024 problems in one attempt, a level of mathematical reasoning usually associated with far larger models. The complete spec sheet from the official model card on Hugging Face looks like this:

How to Run Maple-Preview Locally

The fastest way to try Maple-Preview without installing anything is the live chat demo at chat.deepgrove.ai, which DeepGrove shared on the Hacker News launch thread. For a local setup, the Hugging Face repository carries the official Transformers implementation. That path needs a CUDA environment with Triton and FlashAttention installed, because the custom kernels are what make ternary inference practical on GPUs.

Apple Silicon users get a separate on-device runtime that produced the headline 218 tokens per second on a Mac mini M4 and more than 200 tokens per second on Apple Silicon generally. DeepGrove also maintains a GGUF build at deepgrove/maple-preview-GGUF, so llama.cpp-style loaders such as Ollama and LM Studio can pick it up without the Triton toolchain. If you are building an application around the model, you can wrap it behind a multi-model API fallback system for zero-downtime routing across providers - the same pattern we covered for Cloudflare Workers.

Maple-Preview vs Other On-Device Models

The fairest comparison for Maple-Preview is against other small MoE models designed for local use. DeepGrove claims a five to sixteen times speed advantage over Gemma 4, Qwen3.5 and gpt-oss, and against a 1-bit dense model the gap is even starker. The company notes that 1-bit Bonsai 27B, a Qwen3.6 27B derivative, can take over five minutes to answer a simple question, while Maple-Preview streams at more than 120 tokens per second on an iPhone in DeepGrove's own demonstration. For coding and agent work, though, purpose-built models like Moonshot AI's Kimi K2.7 Code remain the safer pick.

Maple-Preview Limitations: What a Preview Release Means

The model name is literal: this is a preview, not a finished product. DeepGrove's model card warns that the release received minimal post-training for agentic tasks and only small-scale general reinforcement learning, so tool use and long agent chains may not hold together even where pure reasoning does. Independent coverage from AI Weekly makes the same point - the reasoning is impressive, but the model is not yet tuned for agents. Robust agentic behavior needs long-horizon memory and tool grounding, the exact topics we examined in our deep dive on agentic AI and long-horizon memory.

The benchmark caveat matters too. The published scores are DeepGrove's own reporting on its model via a dense output head, and there is no like-for-like third-party table against other models at the same 1B active-parameter budget yet. As with any fresh release, treat the numbers as vendor-reported until independent evals land.

What's Next for DeepGrove's Ternary Line

DeepGrove's announcement on X frames Maple-Preview as "the first step in our journey towards efficient and adaptable intelligence," with agentic training scaling up in future versions. The company's broader bet, in the words of Harj Taggar, is that "local models that are always learning on device will be the catalyst for personal agents we take with us everywhere." Ternary weights are the enabling layer: they make that kind of always-on model small enough and fast enough to live on phones and laptops. Watch for a fully agentic follow-up, plus wider support for macOS, Android and custom silicon in the GGUF and on-device runtimes.

The Bottom Line

Maple-Preview is the strongest proof yet that a 20B-class reasoning model can run comfortably on a laptop or phone. The 78.7% average across LiveCodeBench v6, AIME 2026, HMMT 2026 and GPQA-Diamond, the 5.31 GB checkpoint and the 218 tokens per second on a Mac mini M4 make it a compelling pick for local math, code and general reasoning work. It is not ready for heavy agentic automation, and its benchmark claims are vendor-reported, so independent verification is still pending. For developers who want frontier-style reasoning without a GPU cluster, Maple-Preview is worth downloading today - and the MIT license means you can build on it immediately.

As Harj Taggar, managing partner at Y Combinator, wrote on X: "Personal agents we take with us everywhere" are the destination, and DeepGrove just got a step closer. If you are already running local classifiers, the same setup routine applies to Mistral's new safety classifier - see our guide to running Shieldstral locally for the pattern.