DeepGrove Maple: open 20B ternary MoE, 218 tok/s, 5GB, MIT license
It hits 120 tok/s on an iPhone, solves IMO-level problems, and beats gpt-oss on speed — but ships with minimal agentic tuning and hallucinates facts.
Copy markdown
A 20B reasoner that fits in 5.3GB
Maple-Preview is a 20B-A1B ternary MoE — 24 layers, 256 experts with 8 active — shipping a 5.31GB checkpoint and a 131K-token context. DeepGrove released it under the MIT license, so you can drop it into a commercial app today.
218 tok/s on a Mac mini, 120 on an iPhone
DeepGrove clocks it at 218 tok/s on an M4 Mac mini and about 120 tok/s on an iPhone, claiming 5–16× the decode speed of Gemma 4, Qwen3.5 and gpt-oss at similar quality. The ternary weights are what buy both the speed and the tiny footprint.
It solves IMO problems — and hallucinates confidently
DeepGrove says Maple clears IMO-level math, but warns the preview got "minimal post-training for agentic tasks" and only small-scale RL. Early testers flag aggressive factual hallucination, so keep it grounded with tools or search rather than trusting it on knowledge.
Runs in vLLM, SGLang or Transformers now
The weights are live on Hugging Face with examples for Transformers, vLLM, SGLang and Docker (it needs trust_remote_code=True). There's also a hosted demo at chat.deepgrove.ai if you want to poke it before pulling 5GB.
Build this weekend: an offline reasoning agent
A 5GB reasoner running at 120 tok/s on a phone means copilots that never touch a server — private math tutors, on-device note-reasoners, or edge agents that work in airplane mode. Wire in a search tool to paper over the hallucination gap.