---
title: glm-5.3-flash-w4a8-Ascend-Mixed
canonical_url: "https://www.modelscope.cn/models/japanufo/glm-5.3-flash-w4a8-Ascend-Mixed"
md_url: "https://www.modelscope.cn/models/japanufo/glm-5.3-flash-w4a8-Ascend-Mixed.md"
repository: japanufo/glm-5.3-flash-w4a8-Ascend-Mixed
chinese_name: glm-5.3-flash-w4a8-Ascend-Mixed
last_updated: 2026-10-13
license: "MIT License"
pipeline_tag: image-text-to-text
tasks:
  - image-text-to-text
model_type:
  - glm5_next
architectures:
  - Glm5NextForConditionalGeneration
base_model:
  - ZhipuAI/GLM-5.3-Flash
base_model_relation: quantized
parameters: 218.3B
tensor_type:
  - BF16
  - F32
  - I8
library_name:
  - safetensors
downloads: 40
stars: 3
---

# glm-5.3-flash-w4a8-Ascend-Mixed

> glm-5.3-flash-w4a8-Ascend-Mixed - japanufo 在 ModelScope 开源的模型。本模型名为w4a8版本，实为w8a8版本的精准量化改进，改进规则如下： 13个高敏感层直接保留FP16，不做量化 4个次敏感层用w8a8量化 其它层w4a8量化

japanufo/glm-5.3-flash-w4a8-Ascend-Mixed 是 ModelScope 魔搭社区上的 218.3B 参数image-text-to-text模型，采用 MIT License 许可，基于 ZhipuAI/GLM-5.3-Flash 构建。

- **Repository**: japanufo/glm-5.3-flash-w4a8-Ascend-Mixed
- **License**: MIT License
- **Tasks**: image-text-to-text
- **Parameters**: 218.3B
- **Base model**: ZhipuAI/GLM-5.3-Flash
- **Downloads**: 40
- **Stars**: 3
- **Last updated**: 2026-10-13

Source: https://www.modelscope.cn/models/japanufo/glm-5.3-flash-w4a8-Ascend-Mixed

---

### 本模型名为w4a8版本，实为w8a8版本的精准量化改进，改进规则如下：
- 13个高敏感层直接保留FP16，不做量化
- 4个次敏感层用w8a8量化
- 其它层w4a8量化

### 逐层量化规则

```
Layer  0  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   (dense 入口)
Layer  1  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16
Layer  2  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16
Layer  3  ░░░░░░░░░░░░░░░░░░░░  W8A8   (首个 MoE + MLA)
Layer  4  ····················  W4A8
Layer  5  ····················  W4A8
Layer  6  ····················  W4A8
Layer  7  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer  8  ····················  W4A8
Layer  9  ····················  W4A8
Layer 10  ····················  W4A8
Layer 11  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 12  ░░░░░░░░░░░░░░░░░░░░  W8A8  
Layer 13  ····················  W4A8
Layer 14  ····················  W4A8
Layer 15  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 16  ····················  W4A8
Layer 17  ····················  W4A8
Layer 18  ····················  W4A8
Layer 19  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 20  ····················  W4A8
Layer 21  ····················  W4A8
Layer 22  ····················  W4A8
Layer 23  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 24  ····················  W4A8
Layer 25  ····················  W4A8
Layer 26  ····················  W4A8
Layer 27  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 28  ····················  W4A8
Layer 29  ····················  W4A8
Layer 30  ····················  W4A8
Layer 31  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 32  ····················  W4A8
Layer 33  ····················  W4A8
Layer 34  ····················  W4A8
Layer 35  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 36  ····················  W4A8
Layer 37  ····················  W4A8
Layer 38  ····················  W4A8
Layer 39  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 40  ····················  W4A8
Layer 41  ····················  W4A8
Layer 42  ····················  W4A8
Layer 43  ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   ★ MLA
Layer 44  ░░░░░░░░░░░░░░░░░░░░  W8A8  
Layer 45  ░░░░░░░░░░░░░░░░░░░░  W8A8   ★ MTP/NextN 草稿层（投机解码校验回退，量化不影响最终输出）
非层张量 ▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓▓  FP16   (embedding/output/visual)
```
- 量化后占用约294GB，比w8a8版本节省28GB左右，宝贵空间可获得更大上下文
- 910B 8卡起服成功
- 实测体感持平甚至略好于w8a8版本

#### 您可以通过如下git clone命令，或者ModelScope SDK来下载模型

SDK下载
```bash
#安装ModelScope
pip install modelscope
```
```python
#SDK模型下载
from modelscope import snapshot_download
model_dir = snapshot_download('japanufo/glm-5.3-flash-w4a8-Ascend-Mixed')
```
Git下载
```
#Git模型下载
git clone https://www.modelscope.cn/japanufo/glm-5.3-flash-w4a8-Ascend-Mixed.git
```

<p style="color: lightgrey;">如果您是本模型的贡献者，我们邀请您根据<a href="https://modelscope.cn/docs/ModelScope%E6%A8%A1%E5%9E%8B%E6%8E%A5%E5%85%A5%E6%B5%81%E7%A8%8B%E6%A6%82%E8%A7%88" style="color: lightgrey; text-decoration: underline;">模型贡献文档</a>，及时完善模型卡片内容。</p>