20.2B total / 1.49B active (MoE)
131,072 tokens
MIT
Open-weight preview of DeepGrove's reasoning-focused Maple family, designed from the beginning for fast, memory-efficient inference on consumer devices. Maple-Preview is a 20.2B-total, 1.49B-active ternary-weight Mixture-of-Experts model with 24 layers, 256 routed experts, eight experts selected per token, a 2,048-dimensional hidden state, 16 attention heads, and four key-value heads. Its attention stack repeats three 512-token sliding-window layers followed by one global-attention layer, bounding most KV-cache growth while supporting a 131,072-token context window. The model's weights use ternary values represented as {-alpha, 0, +alpha}; DeepGrove provides the original Safetensors checkpoint, a compressed 5.31 GB 2-bit MLX package, and official TQ1_0 and TQ2_0 GGUF variants under the MIT license. Maple-Preview and Maple-Preview (Flash) share the same underlying transformer weights. The standard mode computes exact logits through the dense output head, while the optional FlashHead artifact clusters the 151,936-token vocabulary into 4,748 clusters and computes exact logits only for the top 512 candidate clusters, with designated special tokens always scored. DeepGrove reports 169 decode tokens per second with the exact head and 218 tokens per second with FlashHead on a 16 GB M4 Mac mini; on an M5 Pro, it reports 359 and 395 tokens per second respectively. Its published reasoning evaluation gives the exact-head model 75.1 on LCBv6, 87.5 on AIME 2026, 78.8 on HMMT 2026, and 73.5 on GPQA-Diamond, averaging 78.7. The FlashHead configuration scores 75.6, 85.8, 80.3, and 69.2 respectively, averaging 77.7. DeepGrove cautions that this preview received minimal agentic post-training and only small-scale general reinforcement learning, so it is focused primarily on raw reasoning and may underperform on agentic tasks ahead of Maple's planned full release.