Model Overview

Description:

The NVIDIA DeepSeek-V4.1-Flash-1017 NVFP4 model is the quantized version of DeepSeek AI's DeepSeek-V4.1-Flash-1017 model, which is an auto-regressive language model that uses an optimized transformer architecture. DeepSeek-V4.1-Flash-1017 is a natively multimodal Mixture-of-Experts (MoE) model with a Causal Encoder-Decoder (CED) architecture, Compressed Sparse Attention 2 (CSA2), and support for contexts of up to one million tokens. For more information, please check here. The NVIDIA DeepSeek-V4.1-Flash-1017 NVFP4 model is quantized with Model Optimizer.

This model is ready for commercial or 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-V4.1-Flash-1017) Model Card from DeepSeek AI.

License/Terms of Use:

MIT

Deployment Geography:

Global

Use Case:

Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG systems, and other AI-powered applications.

Release Date:

Hugging Face 10/30/2026 via https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4

References

NVIDIA Model Optimizer: https://github.com/NVIDIA/Model-Optimizer

Model Architecture:

Architecture Type: Transformers

Network Architecture: DeepSeek-V4.1-Flash-1017 (DeepseekV41ForCausalLM) — Causal Encoder-Decoder Mixture-of-Experts with Compressed Sparse Attention 2

Number of Model Parameters: 552B backbone in total, 8B activated during prefill and 16B during decode, plus 196B Engram conditional memory

This model was developed based on DeepSeek-V4.1-Flash-1017

Input:

Input Type(s): Text, Image

Input Format(s): String, Red, Green, Blue (RGB)

Input Parameters: One-Dimensional (1D), Two-Dimensional (2D)

Other Properties Related to Input: Context length up to 1M

Output:

Output Type(s): Text

Output Format: String

Output Parameters: One-Dimensional (1D): Sequences

Other Properties Related to Output: None

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration:

Supported Runtime Engine(s):

• vLLM (needs MXFP4 Engram support, vllm-project/vllm#61530)

This checkpoint stores the Engram lookup tables as MXFP4, which stock vLLM images cannot load. Use a vLLM build that includes the pull request above. The benchmark results below were obtained with vllm/vllm-openai:v0.31.0 with that pull request's Engram changes applied. SGLang has not been validated with this checkpoint.

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Blackwell

Preferred Operating System(s):

• Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Model Version(s):

The model is quantized with nvidia-modelopt v0.48.0rc0. Source revision: 1bae0e1088993b9b82de08f59de5be5a34e42170.

Training and Evaluation Datasets:

Calibration Dataset:

Link: cnn_dailymail, Nemotron-Post-Training-Dataset-v2

Data Collection Method by dataset: Automated.

Labeling Method by dataset: Automated.

Properties: The cnn_dailymail dataset is an English-language dataset containing just over 300k unique news articles as written by journalists at CNN and the Daily Mail. The Nemotron-Post-Training-Dataset-v2 is a post-training dataset curated by NVIDIA containing multi-turn conversations across diverse topics.

Calibration used 512 samples per dataset (1,024 total), sequence length 512, batch size 4, and the default deterministic sample selection. The Nemotron subsets were stem, chat, math, and code.

Training Dataset:

Data Modality: Undisclosed

Data Collection Method by dataset: Undisclosed

Labeling Method by dataset: Undisclosed

Properties: Undisclosed

Evaluation Dataset:

Datasets: GPQA Diamond, AA-LCR, SciCode, IFBench, MMMU-Pro, Terminal-Bench 2.1

Data Collection Method by dataset: Hybrid: Automated, Manually-Collected

Labeling Method by dataset: Hybrid: Manually-Labeled, Automated

Properties: These benchmarks cover scientific reasoning (GPQA Diamond), long-context recall (AA-LCR), scientific coding (SciCode), instruction following (IFBench), visual understanding (MMMU-Pro), and terminal-based agent tasks (Terminal-Bench 2.1).

Inference:

Acceleration Engines: vLLM

Test Hardware: NVIDIA GB300 (Blackwell)

Post Training Quantization

This model converts the ordinary routed MoE experts from source MXFP4 to NVFP4 weights and activations (W4A4), with group size 16. The converted projections are w1, w2, and w3 for 384 experts across 40 layers. The Engram lookup tables are converted from the source FP8 to MXFP4. Attention, shared experts, vision, MTP, and other excluded components retain their source precision, including MXFP8 where applicable.

All 16,986,931,200 expert weight blocks passed lossless conversion, preserving dequantized weight values with signed-zero canonicalization; the Engram conversion is lossy. Activation scales were calibrated; 24 of 46,080 projection entries used fallback scales. Lossless weight conversion does not imply identical inference outputs.

The checkpoint is approximately 400 GiB (source: approximately 475 GiB) in 48 safetensors shards.

Usage

Deploy with vLLM

This checkpoint needs a vLLM build with MXFP4 Engram support (vllm-project/vllm#61530). The command below is the one used for the reported results on four GB300 GPUs, with the repository ID in place of the local path.

VLLM_ENGINE_READY_TIMEOUT_S=3600 vllm serve nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 \
    --tensor-parallel-size 4 \
    --tokenizer-mode deepseek_v41 \
    --attention-config '{"backend":"FLASHINFER_MLA_SPARSE_DSV41","indexer_kv_dtype":"mxfp4","indexer_sparse_logits":true}' \
    --kv-cache-dtype fp8 \
    --max-num-seqs 128 \
    --engram-config '{"cpu_offload":true}' \
    --reasoning-parser deepseek_v41

Add --enable-auto-tool-choice --tool-call-parser deepseek_v41 for tool calling. Pass chat_template_kwargs: {"thinking": true, "reasoning_effort": "max"} in requests for thinking at maximum effort. Text and image input were validated with this command; speculative decoding (MTP) was not exercised. Validate the final staged package with the intended runtime before deployment.

Evaluation

Accuracy results obtained with vLLM are shown below, in percent.

Precision
GPQA Diamond
AA-LCR
SciCode
IFBench
MMMU-Pro
Terminal-Bench 2.1




MXFP4 (source)
92.77
80.41
57.58
76.33
78.38
81.81




NVFP4 + MXFP4 Engram
92.93
80.19
57.32
76.47
78.73
81.11

Baseline: DeepSeek-V4.1-Flash-1017, a mixed-precision checkpoint with MXFP4 routed experts. Ordinary benchmark requests used temperature=1.0, top_p=1.0, reasoning_effort=max, and max_new_tokens=262144. GPQA used 16 repeats; AA-LCR used 16 repeats in each of two runs; IFBench used 5; SciCode is the mean of 8 runs. Terminal-Bench 2.1 used a separate agentic configuration with eight trajectories per task; the score is the mean of two runs.

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. 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 make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.