README.md

---
title: maccelerate
emoji: ⚡
colorFrom: blue
colorTo: purple
sdk: static
pinned: false
---

# maccelerate

Native MLX model releases for Apple Silicon, with clear provenance and practical runtime guidance.

We publish MLX SafeTensors packs produced by [maccelerate](https://github.com/maccelerate-ai/maccelerate): an open-source converter that reads a GGUF checkpoint's per-tensor quantization allocation and re-quantizes the corresponding clean bf16 source weights for MLX.

## What to expect

- Native MLX model packs for Apple Silicon.
- Per-tensor quantization allocations derived from the named upstream GGUF.
- Reproducibility metadata: source file hashes, allocation summaries, validation results, and shard manifests.
- MTP heads retained when present in the source model.
- Apple-Silicon-focused setup and memory guidance on every model card.

## Current releases

Our first releases are Qwen3.8-27B Unsloth Dynamic v3.0 MLX packs:

- `UD-Q3_K_XL`
- `UD-Q4_K_M`
- `UD-Q4_K_XL`
- `UD-Q5_K_M`
- `UD-Q6_K_M`
- `UD-Q8_K_XL`

They are text-only packs and are intended for [`mlx-serve`](https://github.com/ddalcu/mlx-serve).

## Important compatibility note

These releases are MLX affine re-encodings, not byte-for-byte or numerically identical conversions of llama.cpp K-quant or IQ codebooks. “Same allocation” means the upstream model selected the corresponding bit width for each tensor.

For published Qwen3.8 packs, use `mlx-serve`; stock `mlx-lm` is not supported for this MTP layout.

## Scope

maccelerate currently supports the Qwen3.5 family, including Qwen3.8 27B relatives. Vision towers are not included in the current releases.

Every repository has its own model card with exact size, allocation, source provenance, license, and runtime notes.

## Links

- Conversion code: [maccelerate on GitHub](https://github.com/maccelerate-ai/maccelerate)
- Runtime: [mlx-serve](https://github.com/ddalcu/mlx-serve)
- Models: [maccelerate on Hugging Face](https://huggingface.co/maccelerate/models)