Maple-Preview
#17 in Reasoning ModelsDeepgrove · preview · 2× · last seen Aug 05, 2026
Maple-Preview is an Open Source Reasoning language model from deepgrove with 20 billion parameters, of which approximately 1 billion per Token are active (20B-A1B, Mixture-of-Experts). The weights are natively trained as ternary (2-Bit, values -α/0/+α) rather than quantized post-hoc, making the model extremely memory-efficient with a 5.31 GB Checkpoint and enabling it to run on end devices like Apple-Silicon Macs and iPhones. According to the manufacturer, it achieves SOTA Reasoning performance in its weight class, including on mathematics and coding Benchmarks, but as a pure text model is primarily designed for Reasoning rather than agent capabilities. The model is openly available under MIT license with no per-Token API pricing model.
Features
| Key Benchmark (%) | 87.5% on AIME 2026; 78.8% on HMMT 2026; 75.1% on LiveCodeBench v6; 73.5% on GPQA-Diamond |
| Context Window (Tokens) | 131,072 tokens |
| License | MIT License (open source) |
| Multimodality | None – text-only reasoning model |
| Platform | Hugging Face (BF16 checkpoint), MLX for Apple Silicon (Mac, iPhone), Transformers/CUDA via Triton & FlashAttention, GGUF conversions |
| Price per 1M Tokens | No API price – open weights (MIT license), usable locally/self-hosted |
| Release Date | August 4, 2026 |