README.md · DeepSeek-V4-Flash-0731

1

---

2

license: mit

3

library_name: transformers

4

---

5

# DeepSeek-V4-Flash-0731

6

7

8

9

10

11

<div align="center">

12

  <img src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/logo.svg?raw=true" width="60%" alt="DeepSeek-V4" />

13

</div>

14

<hr>

15

<div align="center" style="line-height: 1;">

16

<a href="https://www.deepseek.com/" target="_blank" style="margin: 2px;">

17

    <img alt="Homepage" src="https://github.com/deepseek-ai/DeepSeek-V2/blob/main/figures/badge.svg?raw=true" style="display: inline-block; vertical-align: middle;"/>

18

</a>

19

<a href="https://chat.deepseek.com/" target="_blank" style="margin: 2px;">

20

    <img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-DeepSeek%20V4-536af5?color=536af5&logoColor=white" style="display: inline-block; vertical-align: middle;"/>

21

</a>

22

</div>

23

<div align="center" style="line-height: 1;">

24

<a href="https://huggingface.co/deepseek-ai" target="_blank" style="margin: 2px;">

25

    <img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-DeepSeek%20AI-ffc107?color=ffc107&logoColor=white" style="display: inline-block; vertical-align: middle;"/>

26

</a>

27

<a href="https://twitter.com/deepseek_ai" target="_blank" style="margin: 2px;">

28

    <img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-deepseek_ai-white?logo=x&logoColor=white" style="display: inline-block; vertical-align: middle;"/>

29

</a>

30

</div>

31

<div align="center" style="line-height: 1;">

32

<a href="LICENSE" style="margin: 2px;">

33

    <img alt="License" src="https://img.shields.io/badge/License-MIT-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>

34

</a>

35

</div>

36

37

<p align="center">

38

<a href="https://arxiv.org/abs/2606.19348"><b>Technical Report</b>👁️</a>

39

</p>

40

41

## Introduction

42

43

**DeepSeek-V4-Flash-0731** is the official release of **DeepSeek-V4-Flash**, superseding the preview version, with substantially enhanced agentic capabilities. It has the same model structure as [DeepSeek-V4-Flash-DSpark](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark), i.e. it comes with a speculative decoding module attached.

44

45

DeepSeek-V4-Flash-0731 outperforms DeepSeek-V4-Pro (Preview) on benchmarks listed below despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available.

46

47

<div align="center">

48

49

50

| :--- | :---: | :---: | :---: | :---: | :---: |

51

| Terminal Bench 2.1 | 82.7 | 61.8 | 72.1 | 81.0 | 85.0 |

52

| NL2Repo | 54.2 | 39.4 | 38.5 | 48.9 | 69.7 |

53

| Cybergym | 76.7 | 38.7 | 52.7 | - | 83.1 |

54

| DeepSWE | 54.4 | 7.3 | 12.8 | 46.2 | 58.0 |

55

| Toolathlon-Verified | 70.3 | 49.7 | 55.9 | 59.9 | 76.2 |

56

| Agents' Last Exam | 25.2 | 15.8 | 16.5 | 23.8 | 25.7 |

57

| AutomationBench Public | 25.1 | 10.8 | 12.8 | 12.9 | 27.2 |

58

| DSBench-FullStack † | 68.7 | 37.0 | 41.8 | 61.8 | 71.6 |

59

| DSBench-Hard † | 59.6 | 25.8 | 31.1 | 54.5 | 71.7 |

60

61

</div>

62

63

Notes:

64

65

1. For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the `max` reasoning effort level with `temperature = 1.0, top_p = 0.95`.

66

2. † DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

67

68

## Chat Template

69

70

This release does not include a Jinja-format chat template. Instead, we provide a dedicated `encoding` folder with Python scripts and test cases demonstrating how to encode messages in OpenAI-compatible format into input strings for the model, and how to parse the model's text output. Please refer to the [`encoding`](encoding/README.md) folder for full documentation.

71

72

The `reasoning_effort` parameter now supports three levels — `low`, `high`, and `max` — which control how much deliberation the model spends before answering.

73

74

A brief example:

75

76

```python

77

from encoding_dsv4 import encode_messages, parse_message_from_completion_text

78

79

messages = [

80

{"role": "user", "content": "hello"},

81

{"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},

82

{"role": "user", "content": "1+1=?"}

83

]

84

85

# messages -> string

86

prompt = encode_messages(messages, thinking_mode="thinking", reasoning_effort="max")

87

88

# string -> tokens

89

import transformers

90

tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-0731")

91

tokens = tokenizer.encode(prompt)

92

```

93

94

## How to Run with vLLM

95

96

DSpark speculative decoding is enabled with a single flag — add --speculative-config with method: dspark to your vLLM launch command:

97

98

`--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'`

99

100

For example, the command below serves the model with vLLM on a single 4×GB300 node.

101

See the [vLLM recipe](https://recipes.vllm.ai/deepseek-ai/DeepSeek-V4-Flash?hardware=b300&features=tool_calling,reasoning) for detailed instructions and other hardware configurations.

102

103

```bash

104

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \

105

--trust-remote-code --kv-cache-dtype fp8 --block-size 256 \

106

--data-parallel-size 4 --enable-expert-parallel \

107

--moe-backend deep_gemm_mega_moe \

108

--attention-config '{"use_fp4_indexer_cache": true}' \

109

--speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

110

```

111

112

## How to Run with SGLang

113

114

Enable DSpark with `--speculative-algorithm DSPARK` and do not set a separate `--speculative-draft-model-path` as the target and draft weights therefore come from the same checkpoint.

115

See the [SGLang cookbook](https://docs.sglang.io/cookbook/autoregressive/DeepSeek/DeepSeek-V4#hw=gb300&variant=flash-official&quant=fp4&strategy=low-latency&nodes=single) for detailed instructions, benchmarks and other hardwares configurations.

116

117

```bash

118

sglang serve \

119

--trust-remote-code \

120

--model-path deepseek-ai/DeepSeek-V4-Flash-0731 \

121

--tp 4 \

122

--moe-runner-backend flashinfer_mxfp4 \

123

--speculative-algorithm DSPARK \

124

--mem-fraction-static 0.90 \

125

--chunked-prefill-size 4096 \

126

--swa-full-tokens-ratio 0.1 \

127

```

128

129

## How to Run Locally

130

131

Please refer to the [inference](inference/README.md) folder for detailed instructions on running DeepSeek-V4 locally, including model weight conversion and interactive chat demos.

132

133

For local deployment, we recommend setting the sampling parameters to `temperature = 1.0`, with `top_p = 0.95` for agentic scenarios and `top_p = 1.0` otherwise. For the `high` and `max` reasoning effort levels, we recommend a maximum output length of **384K** tokens.

134

135

## License

136

137

This repository and the model weights are licensed under the [MIT License](LICENSE).

138

139

## Citation

140

141

```

142

@misc{deepseekai2026deepseekv4,

143

title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},

144

author={DeepSeek-AI},

145

year={2026},

146

}

147

```

148

149

## Contact

150

151

If you have any questions, please raise an issue or contact us at [service@deepseek.com](service@deepseek.com).