Qwen3.8-27B-UD3-Q3_K_XL-MLX

Native MLX affine re-encoding of the official Unsloth Dynamic v3.0 Q3_K_XL per-tensor allocation for Qwen3.8-27B.

This is a 13.63 GB, text-only MLX model with 3.991 effective bits per quantized weight. It retains the model's inline MTP head and is intended for mlx-serve on Apple Silicon.

Before downloading: use this model with mlx-serve. Stock mlx-lm is not a supported runtime for these published Qwen3.8 MTP layouts.

What this model is

The upstream GGUF determines the bit width of each tensor. maccelerate reads that allocation and quantizes the corresponding clean bf16 Qwen weights onto MLX's affine format at the same width.

This model has the same per-tensor bit-width allocation as the source GGUF, re-encoded for MLX. It is not a lossless GGUF conversion and does not claim numerical parity with llama.cpp's K-quant or IQ codebooks.

Claim	Status
Same upstream per-tensor allocation	Yes
Quantized from clean bf16 weights	Yes
Exact numerical parity with the GGUF	No
MTP head	Retained inline; one MTP layer
Vision tower	Omitted; this is a text-only release

Allocation

MLX width	Tensors
8-bit	98
6-bit	7
5-bit	26
4-bit	194
3-bit	157
2-bit	24
Total	506

Widths are not a house preset: each tensor retains the width selected by the upstream Dynamic allocator. The allocation and structural checks for this exact build are recorded in the included manifest.json.

Serving

Apple Silicon requirements

• Apple Silicon Mac running macOS.
• Allow at least 14 GB of disk for the model files, plus working space for the download.
• The model files occupy 13.63 GB. Actual unified-memory needs also depend on the KV cache, context length, and concurrent requests; no minimum-memory configuration is claimed by this card.

Run it with mlx-serve

Install and run mlx-serve according to its README, download this repository locally, then point it at that directory:

mlx-serve --model /path/to/Qwen3.8-27B-UD3-Q3_K_XL-MLX --kv-quant 8

--kv-quant 8 is a reference configuration, not a memory guarantee for every context length or workload. The pack is mixed-width; MTP is present in the checkpoint, so use a runtime configuration that enables speculative decoding if desired.

Runtime compatibility

Use mlx-serve for this release. Stock mlx-lm is not supported for the published Qwen3.8 MTP layout; loading can succeed without providing a supported result.

Reproducibility and generation code

This checkpoint was generated with maccelerate, the open-source converter that reads a Dynamic GGUF's per-tensor allocation and re-encodes the matching clean bf16 model for MLX.

Provenance

Item	Value
Source GGUF	unsloth/Qwen3.8-27B-GGUF, Qwen3.8-27B-UD-Q3_K_XL.gguf
Source GGUF revision	Not recorded in the published manifest
Source GGUF SHA-256	8c2a45ff85e7674ca185ec8eb6cdeab0e617ed9d8018caed0b64380eb2a67a5e
bf16 source	Qwen/Qwen3.8-27B
bf16 source revision	Not recorded in the published manifest
imatrix	imatrix_unsloth.gguf
Converter	maccelerate (revision not recorded in the published manifest)
Output	7 SafeTensors shards; 13.63 GB

The included manifest.json records the source file and hash, allocation, structural validation, and shard hashes for this exact build.

License

The model weights, source GGUF, and conversion code are licensed under Apache-2.0. See the included LICENSE and the upstream model cards for attribution and terms.