DeepGrove ships Maple-Preview, a 20B ternary model built for on-device speed
DeepGrove has released Maple-Preview, an open-source 20B-A1B ternary-weight mixture-of-experts reasoning model, posted to Hacker News as a Show HN by user edwardbzhang. The submission title claims the model runs at 120 tokens per second on an iPhone, but no accessible source documents that figure: the linked deepgrove.ai article returns an HTTP 404 error, and no Wayback Machine snapshot of it exists either.
What can be verified comes from DeepGrove's own Hugging Face model card. It documents inference at 218 tokens per second ("200+ tokens/sec") on a Mac mini M4, which it describes as 5 to 16 times faster than comparable efficient models such as Gemma 4, Qwen3.5, and gpt-oss. No iPhone figure appears anywhere on the model card. An HN commenter, hahahaa, pointed this out directly in the discussion, correcting the submission's framing: the tested device was a Mac mini, not a phone.
Maple-Preview uses a 24-layer, 256-expert configuration with 8 experts active per token, ternary weights, a 5.31 GB checkpoint, and a 131,072-token context window. According to a passage HN commenter kamranjon says is quoted from the (now-inaccessible) DeepGrove article, the company trained the model in ternary format from scratch rather than converting an existing full-precision model down to lower bitwidths, calling that conversion approach "fundamentally the wrong approach." DeepGrove's model card states the model has state-of-the-art reasoning for its weight class, is competitive with larger models, and solves IMO-level math problems; it also shows a benchmark comparison against Qwen 3.5 35B-A3B and a 1-bit Bonsai 27B, but that comparison exists only as an embedded image on the model card, not as extractable text, so the underlying scores could not be independently verified.
DeepGrove flags its own limitations: the preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning, and the company expects it may underperform on agentic benchmarks as a result. Several HN commenters who tried the model, including via an online chat demo at chat.deepgrove.ai, reported it hallucinating confidently on general-knowledge questions that fall outside narrow technical domains. The model is released under the MIT license.
Key facts
- DeepGrove released Maple-Preview, an open-source 20B-A1B ternary-weight MoE reasoning model, posted to Hacker News by user edwardbzhang.
- DeepGrove's own model card documents 218 tok/s on a Mac mini M4, 5 to 16 times faster than Gemma 4, Qwen3.5, and gpt-oss; no source documents the 120 tok/s iPhone claim in the HN submission title, and a commenter directly disputed it.
- Architecture: 24 layers, 256 experts (8 active), ternary weights, 5.31 GB checkpoint, 131,072-token context window.
- The model card claims SOTA reasoning for its weight class and IMO-level math problem solving, and says the model was trained in ternary format from scratch rather than converted from full precision.
- DeepGrove itself warns the preview had minimal agentic post-training and only small-scale general reinforcement learning, and may underperform on agentic benchmarks; it is released under the MIT license.
Why it matters
Ternary-weight models trained natively in low precision, rather than quantized down from a full-precision checkpoint after the fact, are a bet that on-device reasoning models can get faster and smaller without the usual quality loss of post-hoc quantization. Maple-Preview is DeepGrove's attempt at that bet at a 20B-parameter (1B active) scale, and its own numbers show a real speed gain on Apple Silicon over comparable efficient models.
Who it affects
Developers building local or offline AI tools, and Mac users specifically, since the only benchmark DeepGrove documents is a Mac mini M4. Anyone hoping for a confirmed on-device phone benchmark should treat the HN submission's iPhone claim as unverified rather than a DeepGrove-reported figure.
How to use it
Maple-Preview is open-weight and MIT licensed, shipped as a 5.31 GB checkpoint with a 131,072-token context window. DeepGrove also runs an online chat demo at chat.deepgrove.ai that several HN commenters tried directly. No pricing applies since the weights are free to download; DeepGrove has not stated a timeline for a full, non-preview release.
How solid is it
Weak on the specific claim that drives the headline. The deepgrove.ai article the HN post links to 404s, has no Wayback Machine copy, and DeepGrove's own model card documents speed only for a Mac mini M4, not an iPhone. An HN commenter corrected the submission on exactly this point. The model card's benchmark comparison against Qwen 3.5 35B-A3B and Bonsai 27B is an embedded image rather than extractable text, so its scores could not be checked independently.
Risks and caveats
DeepGrove's own model card discloses that the preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning, warning it may underperform on agentic benchmarks. Multiple HN commenters who tested the model reported confident hallucination on general-knowledge questions outside narrow technical topics, consistent with DeepGrove's own caveat.
“Maple-Preview has SOTA reasoning for its weight class and is even competitive with larger models.”
— DeepGrove, Hugging Face model card