DeepGrove's Maple-Preview: a ternary 20B MoE that runs on-device at 200+ tok/s

DeepGrove has released Maple-Preview, an open-source reasoning language model built around two aggressive efficiency bets: a sparse mixture-of-experts design and 1.58-bit ternary weights. On paper it holds 20 billion parameters across 24 layers and 256 experts, but only about 1 billion are active per token (8 experts fire at a time), and the ternary quantization shrinks the full checkpoint to just 5.31 GB. That combination is what lets a nominally 20B model run locally on consumer hardware rather than a datacenter GPU. It ships under an MIT license with a 131K-token context window.

The performance claims are the headline. DeepGrove reports 218 tokens/second on a Mac mini M4 via a custom on-device runtime, and pitches the model as 5–16× faster than comparable-quality models such as Gemma, Qwen3.5, and gpt-oss. On quality, it points to state-of-the-art results for its weight class on reasoning suites (LCBv6, AIME 2026, HMMT 2026, GPQA-D) and claims it can handle IMO-level problems — though the write-up doesn’t publish the actual scores, so those numbers deserve independent verification.

Worth flagging: the Show HN title advertises 120 tok/s on an iPhone, but the published benchmark is the Mac mini M4 figure, with no documented iPhone measurement. DeepGrove also concedes the model has had minimal agentic post-training and only small-scale RL, so it’s a raw reasoning core rather than a polished assistant. The real significance here is the packaging — ternary weights plus fine-grained MoE sparsity as a practical recipe for pushing capable reasoning models onto phones and laptops, keeping inference (and user data) entirely on-device.