The NVIDIA DeepSeek R1 FP4 v2 model is the quantized version of the DeepSeek AI's DeepSeek R1 model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA DeepSeek R1 FP4 model is quantized with TensorRT Model Optimizer.

Compared to nvidia/DeepSeek-R1-FP4, this checkpoint additionally quantizes the wo module in attention layers.

This model is ready for commercial/non-commercial use.

Third-Party Community Consideration

This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party’s requirements for this application and use case; see link to Non-NVIDIA (DeepSeek R1) Model Card.

License/Terms of Use:

MIT

Model Architecture:

Architecture Type: Transformers
Network Architecture: DeepSeek R1

Input:

Input Type(s): Text
Input Format(s): String
Input Parameters: 1D (One Dimensional): Sequences
Other Properties Related to Input: DeepSeek recommends adhering to the following configurations when utilizing the DeepSeek-R1 series models, including benchmarking, to achieve the expected performance: \

Set the temperature within the range of 0.5-0.7 (0.6 is recommended) to prevent endless repetitions or incoherent outputs.
Avoid adding a system prompt; all instructions should be contained within the user prompt.
For mathematical problems, it is advisable to include a directive in your prompt such as: "Please reason step by step, and put your final answer within \boxed{}."
When evaluating model performance, it is recommended to conduct multiple tests and average the results.

Output:

Output Type(s): Text
Output Format: String
Output Parameters: 1D (One Dimensional): Sequences

Software Integration:

Supported Runtime Engine(s):

TensorRT-LLM

Supported Hardware Microarchitecture Compatibility:

NVIDIA Blackwell

Preferred Operating System(s):

Linux

Model Version(s):

** The model is quantized with nvidia-modelopt v0.33.0

Training Dataset:

** Data Collection Method by dataset: Hybrid: Human, Automated
** Labeling Method by dataset: Hybrid: Human, Automated

Testing Dataset:

** Data Collection Method by dataset: Hybrid: Human, Automated
** Labeling Method by dataset: Hybrid: Human, Automated

Evaluation Dataset:

** Data Collection Method by dataset: Hybrid: Human, Automated
** Labeling Method by dataset: Hybrid: Human, Automated

Calibration Datasets:

Calibration Dataset: cnn_dailymail
** Data collection method: Automated.
** Labeling method: Automated.

Inference:

Engine: TensorRT-LLM
Test Hardware: B200

Post Training Quantization

This model was obtained by quantizing the weights and activations of DeepSeek R1 to FP4 data type, ready for inference with TensorRT-LLM. Only the weights and activations of the linear operators within transformer blocks are quantized. This optimization reduces the number of bits per parameter from 8 to 4, reducing the disk size and GPU memory requirements by approximately 1.6x.

Usage

Deploy with TensorRT-LLM

To deploy the quantized FP4 checkpoint with TensorRT-LLM LLM API, follow the sample codes below (you need 8xB200 GPU and TensorRT LLM built from source with the latest main branch):

LLM API sample usage:

from tensorrt_llm import SamplingParams
from tensorrt_llm._torch import LLM
def main():
    prompts = [
        "Hello, my name is",
        "The president of the United States is",
        "The capital of France is",
        "The future of AI is",
    ]
    sampling_params = SamplingParams(max_tokens=32)
    llm = LLM(model="nvidia/DeepSeek-R1-FP4-v2", tensor_parallel_size=8, enable_attention_dp=True)
    outputs = llm.generate(prompts, sampling_params)
    # Print the outputs.
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
# The entry point of the program need to be protected for spawning processes.
if __name__ == '__main__':
    main()

Minimum Latency Server Deployment

If you want to deploy your endpoint to minimize response latency for a single-concurrency or low-concurrency use case, follow the instructions below.

Step 1: Create configuration file (args.yaml)

yaml

moe_backend: TRTLLM
use_cuda_graph: true
speculative_config:
  decoding_type: MTP
  num_nextn_predict_layers: 3
  use_relaxed_acceptance_for_thinking: true
  relaxed_topk: 10
  relaxed_delta: 0.6

Step 2: Start the TensorRT-LLM server

bash

trtllm-serve nvidia/DeepSeek-R1-FP4-v2 \
  --host 0.0.0.0 \
  --port 8000 \
  --backend pytorch \
  --max_batch_size 4 \
  --tp_size 8 \
  --ep_size 2 \
  --max_num_tokens 32768 \
  --trust_remote_code \
  --extra_llm_api_options args.yaml \
  --kv_cache_free_gpu_memory_fraction 0.75

Step 3: Send an example query

bash

curl localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "nvidia/DeepSeek-R1-FP4-v2",
    "messages": [{"role": "user", "content": "Why is NVIDIA a great company?"}],
    "max_tokens": 1024
  }'

Evaluation

The accuracy benchmark results are presented in the table below:

Precision	MMLU Pro	GPQA Diamond	HLE	LiveCodeBench	MATH-500	AIME 2024
FP8 (AA Ref)	84	71	9	62	96	68
FP4	83	71	9	68	96	74

Model Limitations:

The base model was trained on data that contains toxic language and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report security vulnerabilities or NVIDIA AI Concerns here.