---
title: DeepSeek-V4-Flash-FP8
canonical_url: "https://www.modelscope.cn/models/sgl-project/DeepSeek-V4-Flash-FP8"
md_url: "https://www.modelscope.cn/models/sgl-project/DeepSeek-V4-Flash-FP8.md"
repository: sgl-project/DeepSeek-V4-Flash-FP8
last_updated: 2026-07-19
license: mit
model_type:
  - deepseek_v4
architectures:
  - DeepseekV4ForCausalLM
base_model:
  - deepseek-ai/DeepSeek-V4-Flash
base_model_relation: quantized
parameters: 291.0B
tensor_type:
  - BF16
  - I64
  - F32
  - F8_E4M3
library_name:
  - safetensors
  - pytorch
frameworks:
  - pytorch
downloads: 77001
stars: 4
tags:
  - deepseek-v4
  - fp8
  - quantized
---

# DeepSeek-V4-Flash-FP8

> DeepSeek-V4-Flash-FP8 - sgl-project 在 ModelScope 开源的模型。DeepSeek-V4-Flash-FP8

sgl-project/DeepSeek-V4-Flash-FP8 是 ModelScope 魔搭社区上的 291.0B 参数机器学习模型，采用 mit 许可，基于 deepseek-ai/DeepSeek-V4-Flash 构建。

- **Repository**: sgl-project/DeepSeek-V4-Flash-FP8
- **License**: mit
- **Parameters**: 291.0B
- **Base model**: deepseek-ai/DeepSeek-V4-Flash
- **Tags**: deepseek-v4, fp8, quantized
- **Downloads**: 77001
- **Stars**: 4
- **Last updated**: 2026-07-19

Source: https://www.modelscope.cn/models/sgl-project/DeepSeek-V4-Flash-FP8

---

# DeepSeek-V4-Flash-FP8

FP8 re-packaging of [`deepseek-ai/DeepSeek-V4-Flash`](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash).
Model architecture, tokenizer, chat template, and reference `encoding/` are
unchanged from the base repo. No fine-tuning, no retraining — weights only.

# Deployment

SGLang Cookbook: https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4

## License

MIT — see `LICENSE`. Copyright © DeepSeek.