Downloads · 30 days
493
39% of all-time downloads
Qwen3.8-27B-DFlash2-oQ4
NewHorizonGroup/Qwen3.8-27B-DFlash2-oQ4
Qwen3.8-27B-DFlash2-oQ4 is a machine learning model from NewHorizonGroup. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for mlx. The card lists the license as apache-2.0.
z-lab/Qwen3.8-27B-DFlash2 的 MLX 4bit (groupsize=64, affine) 量化版 — 用 mlxvlm.convert 转换, 供 exo dev 分支 (dflash2 支持) 作 coupled drafter 使用。
All-time downloads
1.3K
Public
Parameters
1.9B
1.1 GB on disk
Likes
2
Public
Repo makeup
Click a slice to open those files.
.safetensors1.1 GB · 100%
Parameter types
How the weights are stored.
U321.9B · 100%
Base models
Model card
From the Hugging Face model README
NewHorizonGroup/Qwen3.8-27B-DFlash2-oQ4
z-lab/Qwen3.8-27B-DFlash2 的 MLX 4bit (group_size=64, affine) 量化版 — 用 mlx_vlm.convert 转换, 供 exo dev 分支 (dflash2 支持) 作 coupled drafter 使用。
规格
- 3.85GB bf16 → 1.03GB (4.5 bit/weight 有效)
- 结构不变: 5 层 × hidden 5120, vocab 248320, block_size=8, target_layer_ids=[5,19,33,47,61]
config.json含quantization: {group_size:64, bits:4, mode:affine}与完整dflash_config
验收 (NewHorizon 实测 2026-09-20, Qwen3.8-27B-oQ4 target, 数学题 cap5)
| drafter | rounds | accepted | a_slot |
|---|---|---|---|
| bf16 原版 | 102 | 282/360 | 0.7833 |
| 本 4bit 版 | 101 | 283/362 | 0.7818 |
接受率差异 0.15% (噪声内)。spec 是精确算法, 输出分布仍恒等 target。
用法 (exo dev)
model card 里 coupled_drafter = "NewHorizonGroup/Qwen3.8-27B-DFlash2-oQ4", exo 自动下载并通过
compat 校验 (hidden 5120 / layers 5 / vocab 248320 与 target 一致)。
- Library
- mlx
- Type
- qwen3
- License
- apache-2.0
- Created
- Sep 20, 2026
- Updated
- Sep 20, 2026