NVIDIA Releases NVFP4 Quantized DeepSeek-V4-Pro-0813

September 10, 2026September 20, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

NVIDIA has released “nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark", a fully NVFP4-quantized version of DeepSeek AI’s MoE language model DeepSeek-V4-Pro-0813. This model is available for both commercial and non-commercial use, and operates as a single self-contained checkpoint by including the quantized DSpark draft head. The model was quantized using NVIDIA Model Optimizer and supports inference with SGLang and vLLM.

Specifications

• Architecture: Transformer (Mixture of Experts [MoE] with hybrid attention)
• Parameter Count: 1.65T total parameters, 49B active parameters
• Context Length: Up to 1 million tokens

Performance

This model’s backbone weights are bit-for-bit identical to “nvidia/DeepSeek-V4-Pro-0813-NVFP4", with only the DSpark draft head differing. Consequently, the output results match the target model, and accuracy evaluation results carry over directly. Below is a comparison of pre- and post-quantization evaluation results from the model card.

Additionally, evaluations by the publishers of the original model, “DeepSeek-V4-Pro-0813", report scores of 42.7 / 60.0 on HLE (without/with tools) and 87.9 on Terminal Bench 2.1.

These benchmark results show that the model maintains high performance in advanced reasoning and agent tasks such as GPQA Diamond (measuring PhD-level difficulty in physics, chemistry, and biology) and Terminal-Bench (measuring capability as a terminal operation agent). The difference in scores between the MXFP4-format source version and the NVFP4 format is minimal, keeping accuracy degradation from quantization to a minimum. Furthermore, measurements on MT-Bench and SPEED-Bench demonstrating the effectiveness of DSpark speculative decoding show an overall average acceptance length of around 3.7 to 3.8, achieving efficient speedups via block drafting.

Strengths and Use Cases

The original model, DeepSeek-V4-Pro-0813, excels at advanced reasoning, agentic AI applications, tool-use scenarios, and complex problem-solving tasks such as mathematics, software engineering, and enterprise AI assistants. It features an optimized Transformer architecture employing hybrid attention (Compressed Sparse Attention and Heavily Compressed Attention) and Manifold-Constrained Hyper-Connections, supporting a context length of up to 1 million tokens and reasoning effort levels (low, high, max) for thinking mode control. It also integrates the DSpark speculative decoding module within the same checkpoint, enabling efficient text generation.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 1650.5B parameters

Inference engine support (architecture name matched against each project’s own model registry in its source code, checked 2026-09-19): llama.cpp: registered, vLLM: registered, MLX (mlx-lm): not registered. “Not registered" means the name is absent from that registry today, not that the model cannot run.

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

How to Get It

This model is available on Hugging Face in the safetensors distribution format. It can be accessed and used from the Hugging Face repository page for both commercial and non-commercial purposes.

To deploy using SGLang, run the following command:

python3 -m sglang.launch_server \
    --model-path nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark \
    --tp 8 \
    --trust-remote-code \
    --tool-call-parser deepseekv4 \
    --reasoning-parser deepseek-v4

An example of deploying with vLLM is as follows:

vllm serve nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark \
    --served-model-name nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark \
    --tensor-parallel-size 8 \
    --enable-expert-parallel \
    --reasoning-parser deepseek_v4 \
    --max-model-len 400000 \
    --gpu-memory-utilization 0.9 \
    --kv-cache-dtype fp8 \
    --max-num-batched-tokens 8192 \
    --enable-chunked-prefill \
    --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 128}' \
    --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'

Related Articles

• OpenBMB Releases MiniCPM5-2B: A SOTA 2B On-Device Model
• NVIDIA AIPerf: Benchmarking LLM Inference at Scale
• TensorRT Edge-LLM Accelerates MLPerf Agentic Benchmark

Sources

• https://huggingface.co/nvidia/DeepSeek-V4-Pro-0813-nvfp4-DSpark
• https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813
• https://github.com/NVIDIA/Model-Optimizer

Update History

• 2026-09-20: Rewrote the article from re-collected sources and restored it from draft to published.