txgsync/Maple-Preview-BF16-MLX
0179
1---2license: mit3language: en4library_name: mlx5pipeline_tag: text-generation6base_model: deepgrove/maple-preview7tags:8- mlx9- causal-lm10- mixture-of-experts11- reasoning12- custom-code13- bf1614---15 16# Maple Preview — BF16 MLX conversion17 18This repository contains the full-precision BF16 MLX conversion of [`deepgrove/maple-preview`](https://huggingface.co/deepgrove/maple-preview).19 20## This MLX upload21 22- Approximately 40.5 GB for the complete local model directory (40,448,443,564 bytes).23- Converted directly from the original Hugging Face BF16 safetensors.24- Includes the official Maple model implementation as `maple.py`.25- Requires custom-code loading (`trust_remote_code=True`). In oMLX, enable **Trust Remote Code** for this model.26 27This is an MLX conversion for local inference on Apple Silicon. Please follow the base model's MIT license and usage terms.28 29## Recommended generation settings30 31Maple is a reasoning-heavy model and may spend a substantial part of its response budget thinking. For the OpenAI-compatible API or oMLX UI, start with:32 33```text34temperature: 1.035top_p: 0.9536top_k: 4037min_p: 0.0538repetition_penalty: 1.039max_tokens: 8192 or higher40max context: 131072 tokens (native model limit)41```42 43These sampler values match DeepGrove's Maple `llama.cpp` setup. The model declares a native 131,072-token context window and does not require RoPE/YARN scaling for that window. Actual usable context may be lower on systems constrained by KV-cache memory; do not assume that extending beyond 131,072 tokens is supported.44 45---46 47## Base model description48 49# Maple-Preview50 51**DeepGrove · 2026**52 53Today we introduce Maple-Preview, an open-source 20B-A1B ternary-weight reasoning LLM. Maple-Preview has SOTA reasoning for its weight class and is even competitive with larger models. It solves IMO-level problems and runs at 200+ tokens/sec on a Mac mini M4, 5–16× faster than efficient models like Gemma 4, Qwen3.5, and gpt-oss.54 55- 20B-A1B Model56- 218 tok/s M4 Mac mini57- 5.31 GB Checkpoint58- 131,072 Token context59 6061 62> [!NOTE]63> The included Transformers implementation depends on Triton and FlashAttention64> and is intended for a compatible CUDA environment. The reported Apple Silicon65> result uses a separate on-device runtime.66 67## Architecture68 69Maple-Preview is a 20B-A1B reasoning model designed from the start for efficient on-device inference. It utilizes a 24-layer, 256-expert (8 active) configuration with 3:1 SWA-512:GA attention.70 71## Evaluation72 73On benchmarks, Maple-Preview sets a new point on the Pareto frontier for both memory-to-performance and speed-to-performance, demonstrating its strong reasoning capabilities. However, we note that this preview is focused primarily on raw reasoning and, as such, may underperform on agentic benchmarks. We intend to continue improving general performance through extended training before Maple's full release.74 7576 77Capability comparison using the dense output head across LCBv6, AIME 2026, HMMT 2026, and GPQA-D.78 79## Limitations80 81This preview received minimal post-training for agentic tasks and only82small-scale general reinforcement learning.83 84## License85 86Maple-Preview is released under the [MIT License](LICENSE).87