Maple-Preview
#16 in Reasoning-ModelleDeepgrove Ai · preview · 2× · tolest 05. Aug. 2026
Maple-Preview is an Open Source Reasoning model developed by DeepGrove AI with 20.2 billion parameters (Mixture-of-Experts, approximately 1.49 billion active parameters per Token). It uses a natively trained ternary weight architecture (2-bit, {-α, 0, +α}) instead of post-hoc quantization and is optimized specifically for efficient On-Device Inference on Apple hardware (including Mac Mini M4) and iPhones. According to the developer, the model achieves state-of-the-art Reasoning performance in its weight class on Benchmarks such as AIME 2026, HMMT 2026, LiveCodeBench v6, and GPQA-Diamond, but as a Preview version is primarily designed for pure Reasoning (mathematics, coding) and not yet optimized for agentic tasks. The weights are openly available on Hugging Face under the MIT license.
Features
| Key Benchmark (%) | 87.5% on AIME 2026 (78.7% average across AIME26, HMMT26, LCBv6, GPQA-D) |
| Context Window (Tokens) | 131,072 tokens |
| License | MIT License (open source) |
| Multimodality | Not documented (text-based reasoning model per model card, no multimodal input documented) |
| Platform | Apple Silicon (Mac Mini M4, MacBook Pro M5 Pro, iPhone) via MLX; also Hugging Face/Transformers with CUDA (Triton, FlashAttention) |
| Price per 1M Tokens | No API price – open weights, free to self-host |
| Release Date | August 4, 2026 |