MJPansa/DeepSeek-V4-Flash-0731-NVFP4

DeepSeek V4 Flash 0731-NVFP4

Community NVIDIA-style NVFP4 version

This is a community conversion of the official DeepSeek V4 Flash 0731 checkpoint. DeepSeek's architecture, tokenizer, DSpark/MTP configuration, attention, shared experts, router, head, and other excluded tensors are preserved. This repository is not affiliated with or endorsed by DeepSeek or NVIDIA.

Technical Report 👁️

What was converted

• Calibrated routed-expert inputs across all 43 main transformer layers.
• Used exactly 500,000 calibration tokens across a diverse mixture of SWE and coding trajectories, agentic/tool-use traces, GLM/Kimi traces, CUDA/Nsight material, math/Triton material, and science material.
• Generated routed-expert w1, w2, and w3 input scales from the calibration maxima.
• Converted the routed-expert MXFP4 packed weights losslessly to the NVIDIA-style NVFP4 representation.
• Converted source GS32 block scales to target GS16 block scales by splitting each 32-value source block into two 16-value target blocks with the same exact scale. The packed four-bit weight values are unchanged.

The conversion covers 33,024 routed expert projections. Attention, shared experts, router/head, embeddings, and MTP tensors remain in their official 0731 source representation.

Validation

• 33,024 packed routed weights: byte-identical to the source.
• 33,024 weight-scale tensors: exact GS32→GS16 conversion.
• 66,048 generated global/input scale tensors: validated against the frozen calibration result.
• 6,269 passthrough tensors: byte-identical to the source.
• 48 safetensor shards and 138,365 indexed output tensors validated.

The complete conversion receipt is included as conversion-receipt.json.

Runtime status

The checkpoint was verified to load and generate through vLLM on two NVIDIA DGX Spark systems using tensor parallelism across both machines. An OpenAI-compatible chat smoke test completed successfully. The following is the tested configuration; it favors compatibility and deterministic startup over maximum throughput (--enforce-eager disables CUDA graphs and compilation optimizations).

• vLLM: 0.26.1rc1.dev191+gaef9bd1e9.d20260731
• Tensor parallel size: 2 across 2 nodes
• Maximum model length: 250000 tokens
• Maximum sequences: 8
• Maximum batched tokens / chunked prefill: 4096
• GPU memory utilization: 0.83
• KV cache: FP8, block size 256
• Load format: instanttensor

Start the rank-1 worker first:

vllm serve /path/to/DeepSeek-V4-Flash-0731-NVFP4 \
  --host 0.0.0.0 --port 8000 \
  --trust-remote-code --tokenizer-mode deepseek_v4 \
  --tensor-parallel-size 2 --nnodes 2 --node-rank 1 \
  --master-addr <rank-0-address> --master-port 29513 \
  --headless --distributed-executor-backend mp \
  --load-format instanttensor \
  --max-model-len 250000 --max-num-seqs 8 \
  --max-num-batched-tokens 4096 \
  --gpu-memory-utilization 0.83 --kv-cache-dtype fp8 \
  --block-size 256 --enforce-eager

Then start rank 0 with the same command, changing --node-rank 1 to --node-rank 0 and removing --headless.

This runtime check confirms loading and basic generation compatibility. It is not a claim of production performance certification; the tested eager-mode configuration produced low decode throughput.

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	DeepSeek-V4-Flash-0731	DeepSeek-V4-Flash (Preview)	DeepSeek-V4-Pro (Preview)	GLM-5.2	Opus-4.8
Terminal Bench 2.1	82.7	61.8	72.1	81.0	85.0
NL2Repo	54.2	39.4	38.5	48.9	69.7
Cybergym	76.7	38.7	52.7	-	83.1
DeepSWE	54.4	7.3	12.8	46.2	58.0
Toolathlon-Verified	70.3	49.7	55.9	59.9	76.2
Agents' Last Exam	25.2	15.8	16.5	23.8	25.7
AutomationBench Public	25.1	10.8	12.8	12.9	27.2
DSBench-FullStack †	68.7	37.0	41.8	61.8	71.6
DSBench-Hard †	59.6	25.8	31.1	54.5	71.7

Notes:

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.
2. † 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 = [
    {"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("MJPansa/DeepSeek-V4-Flash-0731-NVFP4")
tokens = tokenizer.encode(prompt)

How to Run with vLLM

For the upstream DeepSeek deployment recipe, DSpark speculative decoding is enabled with a single flag — add --speculative-config with method dspark to your vLLM launch command. For this NVFP4 conversion, use the tested two-node command in the Runtime status section above.

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"}'

For example, the command above serves the upstream model on a single 4×GB300 node. See the vLLM recipe for detailed instructions and other hardware configurations.

How to Run Locally

Please refer to the inference folder for detailed instructions on running DeepSeek-V4 locally, including model weight conversion and interactive chat demos.

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.

License

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

Citation

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

Contact

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