{"id":23000,"date":"2026-08-04T00:54:11","date_gmt":"2026-08-04T00:54:11","guid":{"rendered":"https:\/\/scannn.com\/deepseek-ai-deepseek-v4-flash-0731-%c2%b7-hugging-face\/"},"modified":"2026-08-04T00:54:11","modified_gmt":"2026-08-04T00:54:11","slug":"deepseek-ai-deepseek-v4-flash-0731-%c2%b7-hugging-face","status":"publish","type":"post","link":"https:\/\/scannn.com\/lv\/deepseek-ai-deepseek-v4-flash-0731-%c2%b7-hugging-face\/","title":{"rendered":"deepseek-ai\/DeepSeek-V4-Flash-0731 \u00b7 Hugging Face"},"content":{"rendered":"\n
<\/p>\n
\n \n<\/div>\n\n
\n
\n \"Homepage\"
\n <\/a>
\n
\n \"Chat\"
\n <\/a>\n<\/div>\n
\n
\n \"Hugging
\n <\/a>
\n
\n \"Twitter
\n <\/a>\n<\/div>\n
\n
\n \"License\"
\n <\/a>\n<\/div>\n

\n Technical Report<\/b><\/a>\n<\/p>\n

\n\t<\/p>\n

\t<\/a>
\n\t
\n\t\tIntroduction
\n\t<\/span>
\n<\/h2>\n

DeepSeek-V4-Flash-0731<\/strong> is the official release of DeepSeek-V4-Flash<\/strong>, superseding the preview version, with substantially enhanced agentic capabilities. It has the same model structure as DeepSeek-V4-Flash-DSpark<\/a>, i.e. it comes with a speculative decoding module attached.<\/p>\n

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.<\/p>\n

\n
\n\n\n\n\n\n\n\n\n\n\n\n\n\n
Benchmark<\/th>\nDeepSeek-V4-Flash-0731<\/th>\nDeepSeek-V4-Flash (Preview)<\/th>\nDeepSeek-V4-Pro (Preview)<\/th>\nGLM-5.2<\/th>\nOpus-4.8<\/th>\n<\/tr>\n<\/thead>\n
Terminal Bench 2.1<\/td>\n82.7<\/td>\n61.8<\/td>\n72.1<\/td>\n81.0<\/td>\n85.0<\/td>\n<\/tr>\n
NL2Repo<\/td>\n54.2<\/td>\n39.4<\/td>\n38.5<\/td>\n48.9<\/td>\n69.7<\/td>\n<\/tr>\n
Cybergym<\/td>\n76.7<\/td>\n38.7<\/td>\n52.7<\/td>\n–<\/td>\n83.1<\/td>\n<\/tr>\n
DeepSWE<\/td>\n54.4<\/td>\n7.3<\/td>\n12.8<\/td>\n46.2<\/td>\n58.0<\/td>\n<\/tr>\n
Toolathlon-Verified<\/td>\n70.3<\/td>\n49.7<\/td>\n55.9<\/td>\n59.9<\/td>\n76.2<\/td>\n<\/tr>\n
Agents’ Last Exam<\/td>\n25.2<\/td>\n15.8<\/td>\n16.5<\/td>\n23.8<\/td>\n25.7<\/td>\n<\/tr>\n
AutomationBench Public<\/td>\n25.1<\/td>\n10.8<\/td>\n12.8<\/td>\n12.9<\/td>\n27.2<\/td>\n<\/tr>\n
DSBench-FullStack \u2020<\/td>\n68.7<\/td>\n37.0<\/td>\n41.8<\/td>\n61.8<\/td>\n71.6<\/td>\n<\/tr>\n
DSBench-Hard \u2020<\/td>\n59.6<\/td>\n25.8<\/td>\n31.1<\/td>\n54.5<\/td>\n71.7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<\/div>\n

Notes:<\/p>\n

    \n
  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<\/code> reasoning effort level with temperature = 1.0, top_p = 0.95<\/code>.<\/li>\n
  2. \u2020 DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.<\/li>\n<\/ol>\n

    \n\t<\/p>\n

    \t<\/a>
    \n\t
    \n\t\tChat Template
    \n\t<\/span>
    \n<\/h2>\n

    This release does not include a Jinja-format chat template. Instead, we provide a dedicated encoding<\/code> 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<\/code><\/a> folder for full documentation.<\/p>\n

    The reasoning_effort<\/code> parameter now supports three levels \u2014 low<\/code>, high<\/code>, and max<\/code> \u2014 which control how much deliberation the model spends before answering.<\/p>\n

    A brief example:<\/p>\n

    from<\/span> encoding_dsv4 import<\/span> encode_messages, parse_message_from_completion_text\n\nmessages = [\n    {\"role\"<\/span>: \"user\"<\/span>, \"content\"<\/span>: \"hello\"<\/span>},\n    {\"role\"<\/span>: \"assistant\"<\/span>, \"content\"<\/span>: \"Hello! I am DeepSeek.\"<\/span>, \"reasoning_content\"<\/span>: \"thinking...\"<\/span>},\n    {\"role\"<\/span>: \"user\"<\/span>, \"content\"<\/span>: \"1+1=?\"<\/span>}\n]\n\n\nprompt = encode_messages(messages, thinking_mode=\"thinking\"<\/span>, reasoning_effort=\"max\"<\/span>)\n\n\nimport<\/span> transformers\ntokenizer = transformers.AutoTokenizer.from_pretrained(\"deepseek-ai\/DeepSeek-V4-Flash-0731\"<\/span>)\ntokens = tokenizer.encode(prompt)\n<\/code><\/pre>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tHow to Run with vLLM
    \n\t<\/span>
    \n<\/h2>\n
    DSpark speculative decoding is enabled with a single flag \u2014 add –speculative-config with method: dspark to your vLLM launch command:<\/p>\n
    --speculative-config '{\"method\":\"dspark\",\"num_speculative_tokens\":7,\"draft_sample_method\":\"greedy\"}'<\/code><\/p>\n
    For example, the command below serves the model with vLLM on a single 4\u00d7GB300 node.
    \nSee the     vLLM recipe<\/a> for detailed instructions and other hardware configurations.<\/p>\n
    vllm serve deepseek-ai\/DeepSeek-V4-Flash-0731 \\\n  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \\\n  --data-parallel-size 4 --enable-expert-parallel \\\n  --moe-backend deep_gemm_mega_moe \\\n  --attention-config '{\"use_fp4_indexer_cache\": true}'<\/span> \\\n  --speculative-config '{\"method\":\"dspark\",\"num_speculative_tokens\":7,\"draft_sample_method\":\"greedy\"}'<\/span>\n<\/code><\/pre>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tHow to Run with SGLang
    \n\t<\/span>
    \n<\/h2>\n
    Enable DSpark with --speculative-algorithm DSPARK<\/code> and do not set a separate --speculative-draft-model-path<\/code> as the target and draft weights therefore come from the same checkpoint.
    \nSee the     SGLang cookbook<\/a> for detailed instructions, benchmarks and other hardwares configurations.<\/p>\n
    sglang serve \\\n  --trust-remote-code \\\n  --model-path deepseek-ai\/DeepSeek-V4-Flash-0731 \\\n  --tp 4 \\\n  --moe-runner-backend flashinfer_mxfp4 \\\n  --speculative-algorithm DSPARK \\\n  --mem-fraction-static 0.90 \\\n  --chunked-prefill-size 4096 \\\n  --swa-full-tokens-ratio 0.1 \\\n<\/code><\/pre>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tHow to Run Locally
    \n\t<\/span>
    \n<\/h2>\n
    Please refer to the inference<\/a> folder for detailed instructions on running DeepSeek-V4 locally, including model weight conversion and interactive chat demos.<\/p>\n
    For local deployment, we recommend setting the sampling parameters to temperature = 1.0<\/code>, with top_p = 0.95<\/code> for agentic scenarios and top_p = 1.0<\/code> otherwise. For the high<\/code> and max<\/code> reasoning effort levels, we recommend a maximum output length of 384K<\/strong> tokens.<\/p>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tLicense
    \n\t<\/span>
    \n<\/h2>\n
    This repository and the model weights are licensed under the MIT License<\/a>.<\/p>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tCitation
    \n\t<\/span>
    \n<\/h2>\n
    @misc{deepseekai2026deepseekv4,\n      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},\n      author={DeepSeek-AI},\n      year={2026},\n}\n<\/code><\/pre>\n
    \n\t<\/p>\n
    \t<\/a>
    \n\t
    \n\t\tContact
    \n\t<\/span>
    \n<\/h2>\n
    If you have any questions, please raise an issue or contact us at service@deepseek.com<\/a>.<\/p>\n
    <\/div>\n