DeepGrove open-sources Maple-Preview, a 20B-A1B ternary MoE

A small model release from DeepGrove is more interesting for its shape than for any leaderboard position. Maple-Preview, posted to Hugging Face under MIT, is a 20 billion parameter mixture-of-experts model with only about 1 billion parameters active per token, and the weights are ternary. The whole checkpoint is 5.31 GB.

The architectural bet is unusual enough to note. DeepGrove says it is a 24-layer, 256-expert (8 active) configuration, using a 3:1 ratio of sliding-window to global attention. On an M4 Mac mini, they report 218 tokens per second, which they describe as five to sixteen times faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss. The framing on the model card is that reasoning quality is competitive with heavier models in its class, and that it can solve IMO-level problems.

If that speed number holds up outside their own bench, the practical read is that a real reasoning model, not just a chat model, can now run at interactive speed on hardware people already own. That changes the economics of local coding assistants, math tutors, and long-context research helpers for anyone who was paying per token to a hosted service, and it hands the llama.cpp, Ollama, LM Studio, and Jan ecosystems a permissive checkpoint small enough to fit comfortably in RAM.

The honest caveat is that all of the above is DeepGrove reporting on its own model, and the card itself warns that the preview '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. What the release does not give you in transcribable form is the actual benchmark table across LCBv6, AIME 2026, HMMT 2026, and GPQA-D, nor a like-for-like comparison against other models sitting at the same 1B active-parameter budget. The direction worth watching is whether ternary-weight MoE, packaged this small, becomes a default recipe for on-device reasoning rather than a curiosity.