DeepSeek-V4-Flash-0731

Model Tree & Community Derivatives

DeepSeek-V4-Flash-0731 has a substantial community ecosystem of derived models:

Collection

Part of the DeepSeek-V4 collection (8 items, updated 11 days ago, 848 followers).

Related Models

• DeepSeek-V4-Flash-DSpark — Same model structure with speculative decoding module
• DeepSeek-V4-Pro (Preview) — Pro variant (outperformed by Flash-0731 on benchmarks)

Source: HuggingFace Model Card

Citation

If you use DeepSeek-V4-Flash-0731 in your work, please cite:

@misc{deepseekai2026deepseekv4,
      title={DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence},
      author={DeepSeek-AI},
      year={2026},
}

Technical Report

• arXiv: 2606.19348
• Title: DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence

Source: HuggingFace Model Card

License

This repository and the model weights are licensed under the MIT License.

• Repository: MIT License
• Commercial use: Allowed (per model DB record commercial_use_allowed = True)
• Open source: Yes (is_open_source = True)

The MIT License is one of the most permissive open-source licenses, allowing commercial use, modification, distribution, and private use with minimal restrictions.

Source: HuggingFace Model Card

Recommended Sampling Parameters

General Settings

Agentic vs Non-Agentic Scenarios

Reasoning Effort & Output Length

Evaluation Configuration (Benchmarks)

For Code Agent task evaluation:

• Agent framework: DeepSeek Harness (minimal mode, to be released)
• Reasoning effort: max
• temperature = 1.0, top_p = 0.95

Source: HuggingFace Model Card

Local Deployment Guide

For local deployment, refer to the inference folder for detailed instructions on running DeepSeek-V4 locally, including:

• Model weight conversion
• Interactive chat demos

Recommended Sampling Parameters

Output Length

For high and max reasoning effort levels, a maximum output length of 384K tokens is recommended.

Source: HuggingFace Model Card

SGLang Deployment with DSpark

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

Command

sglang serve \
  --trust-remote-code \
  --model-path deepseek-ai/DeepSeek-V4-Flash-0731 \
  --tp 4 \
  --moe-runner-backend flashinfer_mxfp4 \
  --speculative-algorithm DSPARK \
  --mem-fraction-static 0.90 \
  --chunked-prefill-size 4096 \
  --swa-full-tokens-ratio 0.1

Key Parameters

• --speculative-algorithm DSPARK: Enables DSpark speculative decoding (no separate draft model needed)
• --tp 4: Tensor parallelism across 4 GPUs
• --moe-runner-backend flashinfer_mxfp4: MXFP4 quantization backend
• --mem-fraction-static 0.90: 90% static memory fraction
• --chunked-prefill-size 4096: Chunked prefill for efficiency
• --swa-full-tokens-ratio 0.1: Sliding window attention ratio

Cookbook: SGLang cookbook for DeepSeek-V4

Source: HuggingFace Model Card

vLLM Deployment with DSpark Speculative Decoding

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

Command (single 4×GB300 node)

vllm serve deepseek-ai/DeepSeek-V4-Flash-0731 \
  --trust-remote-code --kv-cache-dtype fp8 --block-size 256 \
  --data-parallel-size 4 --enable-expert-parallel \
  --moe-backend deep_gemm_mega_moe \
  --attention-config '{"use_fp4_indexer_cache": true}' \
  --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

Key Parameters

• --speculative-config: Enables DSpark speculative decoding (7 speculative tokens, greedy draft sampling)
• --kv-cache-dtype fp8: FP8 KV cache for memory efficiency
• --data-parallel-size 4: Data parallelism across 4 GPUs
• --moe-backend deep_gemm_mega_moe: Mega MoE backend for expert computation
• --attention-config: FP4 indexer cache for attention

Recipe: vLLM recipe for DeepSeek-V4-Flash

Source: HuggingFace Model Card

Reasoning Effort Levels

The reasoning_effort parameter supports three levels that control how much deliberation the model spends before answering:

Usage

The reasoning effort is set via the reasoning_effort parameter in the encode_messages() function:

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

Output Length Recommendation

For the high and max reasoning effort levels, a maximum output length of 384K tokens is recommended.

Source: HuggingFace Model Card

Chat Template & Encoding

This release does not include a Jinja-format chat template. Instead, a dedicated encoding folder with Python scripts and test cases is provided, demonstrating how to:

1. Encode messages in OpenAI-compatible format into input strings for the model
2. Parse the model's text output

Please refer to the encoding folder for full documentation.

Example

from encoding_dsv4 import encode_messages, parse_message_from_completion_text

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

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

## string -> tokens
import transformers
tokenizer = transformers.AutoTokenizer.from_pretrained("deepseek-ai/DeepSeek-V4-Flash-0731")
tokens = tokenizer.encode(prompt)

Source: HuggingFace Model Card

Agentic Benchmark Performance

DeepSeek-V4-Flash-0731 benchmark scores across agentic and coding agent benchmarks:

Notes:

1. For Code Agent tasks, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95.
2. † DSBench-FullStack is an internal full-stack development test set; DSBench-Hard is an internal test set of difficult coding-agent problems.

Source: HuggingFace Model Card

Introduction & Key Improvements

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 despite its far smaller activated parameter count, and is broadly competitive with the strongest proprietary models available.

Key highlights:

• Official release superseding the preview version
• Substantially enhanced agentic capabilities
• Same model structure as DeepSeek-V4-Flash-DSpark with speculative decoding module
• Outperforms DeepSeek-V4-Pro (Preview) despite smaller activated parameter count
• Broadly competitive with strongest proprietary models

Source: HuggingFace Model Card