---
title: DeepSeek-V4-Flash-0731-W4AFP8-mt
canonical_url: "https://www.modelscope.ai/models/MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt"
md_url: "https://www.modelscope.ai/models/MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt.md"
repository: MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt
last_updated: 2026-09-23
license: mit
model_type:
  - deepseek_v4
architectures:
  - DeepseekV4ForCausalLM
parameters: 156.0B
tensor_type:
  - I8
  - F32
  - BF16
  - F8_E4M3
  - F8_E8M0
  - I64
library_name:
  - safetensors
downloads: 461
stars: 0
---

# DeepSeek-V4-Flash-0731-W4AFP8-mt

> DeepSeek-V4-Flash-0731-W4AFP8-mt - An open-source model by MooreThreads on ModelScope. DeepSeek-V4-Flash-0731-W4AFP8-mt

MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt is a 156.0B-parameter machine learning model on ModelScope. licensed under mit.

- **Repository**: MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt
- **License**: mit
- **Parameters**: 156.0B
- **Downloads**: 461
- **Stars**: 0
- **Last updated**: 2026-09-23

Source: https://www.modelscope.ai/models/MooreThreads/DeepSeek-V4-Flash-0731-W4AFP8-mt

---

# DeepSeek-V4-Flash-0731-W4AFP8-mt

## 模型介绍

本模型为 **DeepSeek-V4-Flash-0731** MoE 量化版本（W4AFP8 / mt）。

源权重为 packed FP4 的 `DeepSeek-V4-Flash-0731`。结构与 0731 相同（含 DSpark speculative decoding / MTP），仅将 routed / MTP expert 权重量化为 W4AFP8。

## 量化方式

与 `DeepSeek-V4-Flash-W4AFP8` 相同：

- 原始 expert 权重从 packed FP4（e2m1 + e8m0 scale）转换为 signed INT4
- scale 的 group size 从 32 调整为 128
- recipe：`direct-fp4-to-symmetric-int4-group128`
- 仅转换 `layers.*.ffn.experts.*.w{1,2,3}` 与 `mtp.*.ffn.experts.*.w{1,2,3}`
- Attention、shared expert 及其他张量保持原样
- `config.json` 中 `quantization_config.quant_method = w4afp8`，`activation_scheme = dynamic`，`fmt = e4m3`

## 权重约定

- `*.weight`：两个 INT4 打包进一个 `int8`
- `*.weight_scale_inv`：`float32`，shape `[N, K/128]`（示例 `layers.0.ffn.experts.0.w1.weight` 为 `[2048, 2048]` / `[2048, 32]`）
- `*.input_scale`：`bfloat16` 占位
- 量化分片：`model-w4afp8-layer00`–`layer42.safetensors`（43 层 routed）、`model-w4afp8-mtp00`–`mtp02.safetensors`（3 个 MTP）
- 未改张量仍在 `model-00001-of-00048.safetensors` … `model-00048-of-00048.safetensors`