DeepGrove's Maple-Preview: 20B ternary reasoning model runs on iPhone at 120 tok/s
DeepGrove released Maple-Preview, an open-source 20B-A1B ternary mixture-of-experts reasoning model that runs locally on Macs and iPhones at 120–228 tokens per second. Independent tests confirm the speed but flag weak coding and hallucination on niche topics. The MIT licence makes it free for commercial use, signalling a shift toward on-device AI.
Bottom line — Maple-Preview shows a 20B-parameter reasoning model fitting in 5GB and running on a phone, per DeepGrove and independent tests.
Go deeper (7)
- Maple-Preview uses ternary weights (-1, 0, +1) and 256 experts with only 8 active per token, giving 20B total parameters but ~1B active, per DeepGrove.
- DeepGrove claims 218 tok/s on an M4 Mac mini and 120+ tok/s on an iPhone, 5–16× faster than Gemma 4, Qwen3.5, and gpt-oss.
- An independent tester measured 228 tok/s on an M4 Max, confirming the speed, but found it failed on 5 brutal maths problems after 10 minutes each, per the same tester.
- The model hallucinated on a niche etymology question, confidently giving a wrong answer, while Qwen 3.6 handled it perfectly, per a Hacker News commenter.
- DeepGrove warns the preview received minimal post-training for agentic tasks, so tool use and long agent chains may not hold, per the model card.
- Maple-Preview ships under MIT licence, free for commercial use, and works with any OpenAI-compatible API endpoint, per the release.
- It lands amid other on-device reasoning work, like Swiftlet running a 4-bit Qwen3-Next 80B in 4.3GB on a Mac last week, per aiweekly.co.