The archive · AI & Models · Technical decision · 2026
DeepGrove bets ternary-weight MoE can run reasoning AI on iPhones
Open-source 20B-A1B ternary MoE (5.31GB) runs ~120-218 tok/s on Apple silicon; HN front page 173 pts, 53 comments.
DeepGrove
What the business is
DeepGrove is a US AI research lab building frontier-quality intelligence for any device; Maple-Preview is its open-source 20B-A1B ternary-weight MoE reasoning model (24 layers, 256 experts, 8 active, MIT) with an online chat demo.
How it started
DeepGrove released Maple-Preview on Hugging Face under MIT in early August 2026, framed as a reasoning LLM designed from the start for efficient on-device inference: 24 layers, 256 experts (8 active), a 3:1 sliding-window-to-global attention ratio and 131,072-token context.
What happened
DeepGrove reported 218 tokens/s on an M4 Mac mini (five to sixteen times faster than Gemma 4, Qwen3.5 and gpt-oss), about 120-127 tok/s on iPhone, and an IMO 2024 P1 solve of 7/7 at 281.5 tok/s on an M5 Pro MacBook Pro. The Show HN drew 173 points and 53 comments; commenters debated hallucination at two-bit sizes, and the llama.cpp project opened an issue requesting Maple support. The model card itself warns the preview received minimal post-training for agentic tasks.
How it ended up
Still live as an open-source preview; DeepGrove is iterating toward a system that adapts models per user and conversation.
Background
DeepGrove is a US AI research lab betting that frontier-quality intelligence can run on any device if models are designed for it from the start. Its argument: quantizing an existing large model is the wrong way to make local AI work, because it degrades quality while still eating memory and bandwidth.
Maple-Preview, released under MIT on Hugging Face in early August 2026, is the evidence: a 20B-parameter MoE with only ~1.49B active per token, weights constrained to ternary values {-1, 0, 1}, 24 layers, 256 experts (8 active), a 3:1 sliding-window-to-global attention ratio, 131,072-token context and a 5.31GB checkpoint.
DeepGrove reports 218 tok/s on an M4 Mac mini, roughly 120-127 tok/s on iPhone, and an IMO 2024 P1 solve of 7/7 at 281.5 tok/s on an M5 Pro MacBook Pro — five to sixteen times faster than efficient models like Gemma 4, Qwen3.5 and gpt-oss. The model card warns it received minimal post-training for agentic tasks.
The Show HN on 2026-08-04 hit the front page with 173 points and 53 comments, and the llama.cpp project opened an issue requesting Maple support. As of September 2026 the model is a preview: DeepGrove says it is iterating and building a system that adapts models per user and conversation.
What has to be true
- Ternary weights invert the local-AI tradeoff: if matmuls become additions, memory and bandwidth collapse without shrinking the model's reasoning.
- Open-sourcing under MIT turned a lab claim into ecosystem infrastructure — llama.cpp support requests and community benchmarks followed within days.
- Shipping numbers on hardware people own (M4 Mac mini, iPhone) made the speed claim concrete, unlike a leaderboard table.
- The bet is structural, not incremental: if 5GB can hold a reasoning model, the economics of coding assistants, math tutors and long-context local tools change.
What can be applied
On-device AI's lever is bits per weight, not parameter count: a 5.31GB ternary MoE that runs at interactive speed earns instant ecosystem attention.
Aftermath
As of 2026-09-02, Maple-Preview is live as an open-source preview: an MIT-licensed checkpoint on Hugging Face with GGUF conversions, community runtimes and an online chat demo. DeepGrove says it will keep improving the model and build a system that adapts to each conversation and user; its stated direction is frontier intelligence on any device, so the release is the opening move of a product line rather than a finished product.
Sources
- Show HN: Maple-Preview – Ternary 20B MoE running at 120 tok/s on a iPhone
- DeepGrove open-sources Maple-Preview, a 20B-A1B ternary MoE
- 苹果 iPhone 上最快本地 AI 模型：Maple-Preview-20B-A1B 登场，可达 127 tokens/s
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card