Qwen3.8-27B-UD3-Q8_K_XL-MLX

Native MLX affine re-encoding of the official Unsloth Dynamic v3.0 Q8_K_XL per-tensor allocation for Qwen3.8-27B.

This is a 31.44 GB, text-only MLX model with 8.500 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, not stock mlx-lm. Stock mlx-lm 0.31.3 can load this converted Qwen3.8 layout but produces corrupted output. See Runtime compatibility.

Genuine Unsloth Dynamic v3.0: maccelerate reads the widths from the official unsloth/Qwen3.8-27B-GGUF UD-Q8_K_XL checkpoint's own ggml type table. It does not use an older Dynamic 2.0 class map.

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 therefore 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 codebooks.

Claim	Status
Same upstream per-tensor allocation	Yes
Quantized from clean bf16 weights	Yes
Exact numerical parity with the GGUF	No

Allocation

MLX width	Tensors
8-bit	453
Total	453

Unlike the Q4–Q6 variants, this source allocation maps entirely to 8-bit MLX affine tensors. Norms, biases, convolution weights, and SSM state remain bf16. The vision tower is deliberately omitted.

Weight class	Parameters	On disk
MLP gate + up	11.41B	12.12 GB
MLP down	5.70B	6.06 GB
GDN in-proj QKV	2.52B	2.67 GB
GDN in-proj Z	1.51B	1.60 GB
GDN out-proj	1.51B	1.60 GB
Token embeddings	1.27B	1.35 GB
Attention Q/K/V/O	0.50B	0.53 GB
MTP head	0.30B	0.32 GB
GDN A/B gates	0.02B	0.03 GB
Total quantized	24.75B	26.29 GB

The MTP head is stored inline: the index contains 23 language_model.mtp.* tensors and the config declares one MTP layer.

Serving

Apple Silicon requirements

• Apple Silicon Mac running macOS.
• Allow at least 32 GB of disk for the model files, plus working space for the download.
• No minimum unified-memory configuration has been validated for this variant. Its model files alone are 31.44 GB; generation, KV cache, context length, and concurrent requests need additional memory.

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-Q8_K_XL-MLX --kv-quant 8

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

Runtime compatibility

This release is not supported by stock mlx-lm. With mlx-lm 0.31.3, loading can succeed but its greedy output is corrupted: it applies Qwen3.8's norm transformation a second time because it treats the presence of MTP weights as evidence that every trunk norm needs adjustment. Use mlx-serve for this model.

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. Its repository contains the conversion workflow, validation code, and the mlx-lm compatibility finding.

Provenance

Item	Value
Source GGUF	unsloth/Qwen3.8-27B-GGUF, Qwen3.8-27B-UD-Q8_K_XL.gguf
Source GGUF SHA-256	af36ecb6b5db1407953345b746c14ac93f0657dda413910b4348683a2d990377
bf16 source	Qwen/Qwen3.8-27B
imatrix	imatrix_unsloth.gguf
Output	15 SafeTensors shards; 31.44 GB

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