txgsync/Maple-Preview-oQ4e

Maple Preview — oQ4e MLX quantization

Fix note: Initial testing found incoherence and repetitive thinking loops because Maple's routed expert tensors were not receiving their corresponding iMatrix weighting; this replacement applies the missing expert/fused-attention iMatrix mapping and has been runtime-tested.

This repository contains an enhanced mixed-precision oQ4e MLX quantization of deepgrove/maple-preview, converted from the BF16 safetensors with oMLX.

This MLX quant upload

• Approximately 11.8 GB for the complete local model directory (11,786,973,625 bytes).
• Uses the same 1,024-sample, 512-token enhanced iMatrix calibration as the oQ8e conversion.
• 18,432 ordinary tensors are 4-bit, with sensitivity-selected 5-, 6-, and 8-bit upgrades; no quantized tensor is below 4-bit.
• The 24 MoE router matrices and all normalization weights remain BF16.
• lm_head and model.word_embeddings are 8-bit.
• Maple's fused Q/K/V projections use shape-compatible per-layer formats and have been runtime-tested through oMLX.
• Requires custom-code loading (trust_remote_code=True). In oMLX, enable Trust Remote Code for this model.

This is an MLX conversion for local inference on Apple Silicon. Please follow the base model's MIT license and usage terms.

Recommended generation settings

Maple 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:

temperature:         1.0
top_p:               0.95
top_k:               40
min_p:               0.05
repetition_penalty:  1.0
max_tokens:          8192 or higher
max context:         131072 tokens (native model limit)

These 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.

---

Base model description

Maple-Preview

DeepGrove · 2026

Today 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.

• 20B-A1B Model
• 218 tok/s M4 Mac mini
• 5.31 GB Checkpoint
• 131,072 Token context

The included Transformers implementation depends on Triton and FlashAttention and is intended for a compatible CUDA environment. The reported Apple Silicon result uses a separate on-device runtime.

Architecture

Maple-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.

Evaluation

On 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.

Capability comparison using the dense output head across LCBv6, AIME 2026, HMMT 2026, and GPQA-D.

Limitations

This preview received minimal post-training for agentic tasks and only small-scale general reinforcement learning.

License

Maple-Preview is released under the MIT License.