deepseek-ai/DeepSeek-V4-Flash-0731 · Hugging Face

Log In<br>Sign Up

","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"eos_token":{"__type":"AddedToken","content":"","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"pad_token":{"__type":"AddedToken","content":"","lstrip":false,"normalized":true,"rstrip":false,"single_word":false},"unk_token":null}},"createdAt":"2026-07-31T07:30:24.000Z","discussionsDisabled":false,"discussionsSorting":"recently-created","downloads":0,"downloadsAllTime":0,"id":"deepseek-ai/DeepSeek-V4-Flash-0731","isLikedByUser":false,"availableInferenceProviders":[],"showHuggingChatEntry":false,"inference":"","lastModified":"2026-07-31T12:02:14.000Z","likes":60,"pipeline_tag":"text-generation","library_name":"transformers","librariesOther":[],"trackDownloads":true,"model-index":null,"private":false,"repoType":"model","gated":false,"tags":["transformers","safetensors","deepseek_v4","text-generation","arxiv:2606.19348","license:mit","endpoints_compatible","8-bit","fp8","region:us"],"tag_objs":[{"id":"text-generation","label":"Text Generation","type":"pipeline_tag","subType":"nlp"},{"id":"transformers","label":"Transformers","type":"library"},{"id":"safetensors","label":"Safetensors","type":"library"},{"id":"deepseek_v4","label":"deepseek_v4","type":"other","clickable":true},{"id":"endpoints_compatible","label":"Inference Endpoints","type":"other","clickable":true},{"id":"8-bit","label":"8-bit precision","type":"other","clickable":true},{"id":"fp8","label":"fp8","type":"other","clickable":true},{"id":"arxiv:2606.19348","label":"arxiv:2606.19348","type":"arxiv","extra":{"paperTitle":"DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence"}},{"id":"license:mit","label":"mit","type":"license"},{"type":"region","label":"🇺🇸 Region: US","id":"region:us"}],"transformersInfo":{"auto_model":"AutoModelForCausalLM","pipeline_tag":"text-generation","processor":"AutoTokenizer"},"widgetData":[{"text":"My name is Julien and I like to"},{"text":"I like traveling by train because"},{"text":"Paris is an amazing place to visit,"},{"text":"Once upon a time,"}],"safetensors":{"parameters":{"BF16":1483567488,"I64":2327040,"F32":37741630,"F8_E4M3":6304038912,"I8":296352743424},"total":304180418494,"sharded":true,"totalFileSize":166892138207},"hasBlockedOids":false,"region":"us","isQuantized":false,"licenseFilePath":"LICENSE"},"discussionsStats":{"closed":0,"open":1,"total":1},"query":{},"inferenceContextData":{"billableEntities":[],"entityName2Providers":{}},"hasQuantizations":false,"copyToBucketNamespaces":[]}">

DeepSeek-V4-Flash-0731

Technical Report 👁️

Introduction

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, i.e. it comes with a speculative decoding module attached.

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.

Benchmark<br>DeepSeek-V4-Flash-0731<br>DeepSeek-V4-Flash (Preview)<br>DeepSeek-V4-Pro (Preview)<br>GLM-5.2<br>Opus-4.8

Terminal Bench 2.1<br>82.7<br>61.8<br>72.1<br>81.0<br>85.0

NL2Repo<br>54.2<br>39.4<br>38.5<br>48.9<br>69.7

Cybergym<br>76.7<br>38.7<br>52.7<br>83.1

DeepSWE<br>54.4<br>7.3<br>12.8<br>46.2<br>58.0

Toolathlon-Verified<br>70.3<br>49.7<br>55.9<br>59.9<br>76.2

Agents' Last Exam<br>25.2<br>15.8<br>16.5<br>23.8<br>25.7

AutomationBench Public<br>25.1<br>10.8<br>12.8<br>12.9<br>27.2

DSBench-FullStack †<br>68.7<br>37.0<br>41.8<br>61.8<br>71.6

DSBench-Hard †<br>59.6<br>25.8<br>31.1<br>54.5<br>71.7

Notes:

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.

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

Chat Template

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 folder for full documentation.

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

A brief example:

from encoding_dsv4 import encode_messages, parse_message_from_completion_text

messages = [<br>{"role": "user", "content": "hello"},<br>{"role": "assistant", "content": "Hello! I am DeepSeek.", "reasoning_content": "thinking..."},<br>{"role": "user", "content": "1+1=?"}

# messages -> string<br>prompt = encode_messages(messages, thinking_mode="thinking", reasoning_effort="max")

# string -> tokens<br>import transformers<br>tokenizer =...