txgsync/Maple-Preview-oQ8e

1---2license: mit3language: en4library_name: mlx5pipeline_tag: text-generation6base_model: deepgrove/maple-preview7base_model_relation: quantized8tags:9- mlx10- causal-lm11- mixture-of-experts12- reasoning13- custom-code14- quantized15- oq8e16---17 18# Maple Preview — oQ8e MLX quantization19 20This repository contains an enhanced oQ8e MLX quantization of [`deepgrove/maple-preview`](https://huggingface.co/deepgrove/maple-preview), converted from the BF16 safetensors with oMLX.21 22## This MLX quant upload23 24- Approximately 21.5 GB for the complete local model directory (21,497,386,683 bytes).25- Uses a 1,024-sample, 512-token enhanced iMatrix calibration.26- All quantized tensors are 8-bit.27- The 24 MoE router matrices and all normalization weights remain BF16.28- `lm_head` and `model.word_embeddings` are 8-bit.29- Requires custom-code loading (`trust_remote_code=True`). In oMLX, enable **Trust Remote Code** for this model.30 31This is an MLX conversion for local inference on Apple Silicon. Please follow the base model's MIT license and usage terms.32 33## Recommended generation settings34 35Maple 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:36 37```text38temperature:         1.039top_p:               0.9540top_k:               4041min_p:               0.0542repetition_penalty:  1.043max_tokens:          8192 or higher44max context:         131072 tokens (native model limit)45```46 47These 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.48 49---50 51## Base model description52 53# Maple-Preview54 55**DeepGrove · 2026**56 57Today 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.58 59- 20B-A1B Model60- 218 tok/s M4 Mac mini61- 5.31 GB Checkpoint62- 131,072 Token context63 64![Maple-Preview speed and performance frontier](assets/01-speed-frontier.png)65 66> [!NOTE]67> The included Transformers implementation depends on Triton and FlashAttention68> and is intended for a compatible CUDA environment. The reported Apple Silicon69> result uses a separate on-device runtime.70 71## Architecture72 73Maple-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.74 75## Evaluation76 77On 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.78 79![Benchmark score comparison](assets/05-benchmark-scores-table.png)80 81Capability comparison using the dense output head across LCBv6, AIME 2026, HMMT 2026, and GPQA-D.82 83## Limitations84 85This preview received minimal post-training for agentic tasks and only86small-scale general reinforcement learning.87 88## License89 90Maple-Preview is released under the [MIT License](LICENSE).91