Downloads · 30 days
167
15% of all-time downloads
adrianmurray/Qwen3.8-27B-MLX-6bit
Qwen3.8-27B-MLX-6bit is a text generation model from adrianmurray. Use it when you need the model to write or continue text. It is set up for mlx. The card lists the license as apache-2.0.
This is a modified, 6-bit MLX quantization of Alibaba Cloud's official Qwen/Qwen3.8-27B checkpoint. It is intended for local inference on Apple Silicon and can be paired with the separately distributed native-MTP arti…
Downloads · 30 days
167
15% of all-time downloads
All-time downloads
1.1K
Public
Parameters
26.9B
21.9 GB on disk
Likes
2
Public
Click a slice to open those files.
.safetensors21.9 GB · 100%
How the weights are stored.
U3226.9B · 100%
From the Hugging Face model README
This is a modified, 6-bit MLX quantization of Alibaba Cloud's official
Qwen/Qwen3.8-27B checkpoint. It is intended for local inference on Apple
Silicon and can be paired with the separately distributed native-MTP artifact
from the same source revision.
Qwen/Qwen3.8-27B1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0The internal Transformers/MLX architecture identifiers retain the
qwen3_5 name because Qwen3.8 uses the compatible architectural foundation.
Those implementation identifiers do not change the checkpoint identity.
Using MLX LM 0.31.3 and MLX 0.32.0:
mlx_lm.convert \
--hf-path /path/to/Qwen3.8-27B \
--mlx-path /path/to/Qwen3.8-27B-MLX-6bit \
--quantize --q-bits 6 --q-group-size 64 --q-mode affine
The source directory must correspond to the exact base revision above. Verify
the converted files against SHA256SUMS before publishing or loading them.
qwen-prime-runtime configure \
--target /path/to/Qwen3.8-27B-MLX-6bit \
--draft /path/to/Qwen3.8-27B-MTP-MLX-6bit
qwen-prime-runtime doctor
qwen-prime-runtime serve
No throughput is guaranteed. On the development M4 Max, a warm 256-token coding request measured approximately 26 server tokens/second with 53.9% draft acceptance. A direct block-size sweep measured approximately 28 tokens/second at block size four. Results vary with hardware, prompt length, cache state, generation length, and acceptance ratio.
The original Qwen3.8-27B model is Copyright 2026 Alibaba Cloud and is licensed
under Apache License 2.0. This directory includes the original license. The
weights and metadata in this repository were modified by conversion and
quantization for MLX; see NOTICE and quantization_provenance.json.
The Apache license does not grant trademark rights. This is an independent derivative and is not endorsed by Alibaba Cloud or the Qwen team.