---
title: Deepseek-v4-flash-w4a8
canonical_url: "https://www.modelscope.cn/models/ml1019/Deepseek-v4-flash-w4a8"
md_url: "https://www.modelscope.cn/models/ml1019/Deepseek-v4-flash-w4a8.md"
repository: ml1019/Deepseek-v4-flash-w4a8
last_updated: 2026-06-23
license: "Apache License 2.0"
pipeline_tag: text-generation
tasks:
  - text-generation
model_type:
  - deepseek_v4
architectures:
  - DeepseekV4ForCausalLM
base_model:
  - deepseek-ai/DeepSeek-V4-Flash
base_model_relation: quantized
library_name:
  - safetensors
  - pytorch
frameworks:
  - Pytorch
downloads: 156
stars: 2
---

# Deepseek-v4-flash-w4a8

> Deepseek-v4-flash-w4a8 - ml1019 在 ModelScope 开源的模型。You are viewing the default Readme template as no detailed model-card was provided by the model’s contributors. You can access the model files in the "Files and versions" tab. Model files may be downloaded…

ml1019/Deepseek-v4-flash-w4a8 是 ModelScope 魔搭社区上的text-generation模型，采用 Apache License 2.0 许可，基于 deepseek-ai/DeepSeek-V4-Flash 构建。

- **Repository**: ml1019/Deepseek-v4-flash-w4a8
- **License**: Apache License 2.0
- **Tasks**: text-generation
- **Base model**: deepseek-ai/DeepSeek-V4-Flash
- **Downloads**: 156
- **Stars**: 2
- **Last updated**: 2026-06-23

Source: https://www.modelscope.cn/models/ml1019/Deepseek-v4-flash-w4a8

---

### You are viewing the default Readme template as no detailed model-card was provided by the model’s contributors. You can access the model files in the "Files and versions" tab.
#### Model files may be downloaded with ModelScope SDK or through git clone directly.

Download with ModelScope’s Python SDK
```bash
#Install ModelScope
pip install modelscope
```
```python
#Download with ModelScope’s Python SDK
from modelscope import snapshot_download
model_dir = snapshot_download('ml1019/Deepseek-v4-flash-w4a8')
```
Download with Git clone
```
git clone https://www.modelscope.cn/ml1019/Deepseek-v4-flash-w4a8.git
```
### 精度验证

精度 summary 来源：

```text
/mnt/nvme/ml1019/benchmark/ais_bench/datasets/outputs/default/*/summary/summary_*.md
```

default GPQA 有效结果共 `13` 次。

最佳结果来源：

```text
/mnt/nvme/ml1019/benchmark/ais_bench/datasets/outputs/default/20260524_183921/summary/summary_20260524_183921.md
```

结果：

| dataset | version | metric | mode | 最好分数 | 平均分数 | 接近 71.2 的次数 |
|---|---|---|---|---:|---:|---:|
| gpqadataset | - | accuracy | gen | 71.72 | 65.97 | 2 / 13 |

“接近 71.2” 按与 `71.2` 的绝对差不超过 `1.0` 分统计，命中的运行如下：

| run | score | 与 71.2 的绝对差 |
|---|---:|---:|
| 20260524_182949 | 71.21 | 0.01 |
| 20260524_183921 | 71.72 | 0.52 |

基于多次 `ais_bench` GPQA 测试，W4A8 DS4 最好结果达到 `71.72`，另一次结果为 `71.21`，与官方 W8A8-MTP GPQA 参考精度 `71.2` 基本一致。考虑到官方也说明 `ais_bench` 精度存在波动，当前 W4A8 DS4 量化产物在已测试服务配置下可认为在 GPQA 上接近无损。该结论目前主要基于 GPQA；若要扩展为更全面的接近无损结论，仍建议补充 MMLU-Pro 等数据集复测。

### 官方量化报告参考

官方 W8A8-MTP 量化报告使用 `ais_bench` 做精度验证：

| 模型名 | 量化格式 | 数据集 | 测试精度 % | 官方精度 % | 备注 |
|---|---|---|---:|---:|---|
| ml1019/Deepseek-v4-flash-w4a8 | w8a8 | gpqa | 71.21 | 71.2 | V4-Flash Non-Think |
| ml1019/Deepseek-v4-flash-w4a8 | w8a8 | mmlupro | 82.85 | 83.0 | V4-Flash Non-Think |
| ml1019/Deepseek-v4-flash-w4a8 | w8a8 | mmlupro | 85.86 | 86.2 | V4-Flash Max |

使用 `ais_bench` 时，Non-Think 模式 `max_out_len = 65536`，Max 模式 `max_out_len = 131072`。精度存在波动，建议多次测试。
