DeepGrove has published Maple-Preview, a 20B-parameter ternary-weight Mixture-of-Experts model designed specifically for on-device reasoning, achieving remarkable performance metrics on consumer hardware.
The model runs at 218 tokens per second on an M4 Mac mini — representing a 5-16x speedup compared to Gemma 4, Qwen3.5, and gpt-oss running on equivalent hardware. The 5.31 GB checkpoint enables deployment on devices previously considered impractical for reasoning workloads.
Maple-Preview uses a 128K context window and features a 24-layer architecture with 256 experts, of which 8 remain active during inference. The model is released under the MIT license, making it available for both commercial and research applications.
The benchmark coverage includes LCBv6, AIME 2026, HMMT 2026, and GPQA-D — representing reasoning-focused evaluation suites that test mathematical, scientific, and coding capabilities. This positions Maple-Preview as a practical solution for developers seeking reasoning capabilities without cloud API dependencies.
The ternary-weight approach is significant: by constraining weights to just three values (-1, 0, +1), the model achieves dramatic compression while maintaining competitive performance. This technique reduces memory footprint and enables faster inference through simplified computations.
For enterprise deployments, the on-device capability addresses key concerns around data privacy and latency. Reasoning workloads that previously required cloud API calls can now execute locally, reducing both costs and data exposure risks.
The release reflects a broader trend toward efficient on-device AI. As model optimization techniques mature, the boundary between cloud and edge AI continues to shift. Maple-Preview demonstrates that frontier-level reasoning capabilities may soon be accessible entirely on consumer devices.
Developers can access the model checkpoint through DeepGrove’s repository, with documentation covering deployment across different Apple Silicon configurations. The MIT license removes traditional barriers to commercial adoption.