NVIDIA · Released, not on our provider

DeepSeek V4 Pro 0813 NVFP4 (2026): what NVIDIA’s 4-bit checkpoint is, what it needs, its license and accuracy

DeepSeek V4 Pro 0813 NVFP4 is NVIDIA’s quantized checkpoint of DeepSeek’s V4 Pro 0813 model, a 1.65T-parameter mixture-of-experts with 49B active parameters and 4-bit NVFP4 experts for NVIDIA Blackwell GPUs, published on Hugging Face on August 27, 2026 under the MIT license, with NVIDIA’s own table showing accuracy close to the original.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with DeepSeek V4.1 Flash, not DeepSeek V4 Pro 0813 NVFP4.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
NVIDIA
Type
Text
Released
2026-08-27
OpenRouter id
None

Checked

DeepSeek V4 Pro 0813 NVFP4: facts, with sources

  • NVIDIA’s card lists the release date of the NVFP4 checkpoint as August 27, 2026 on Hugging Face, with nvidia/DeepSeek-V4-Pro-0813-NVFP4 as the address. The card says it was quantized with nvidia-modelopt version 0.47.0rc1.Source: huggingface.co · checked 2026-11-19
  • The base model, DeepSeek-V4-Pro-0813, was published by DeepSeek earlier: Hugging Face’s API shows its repository was created on August 13, 2026. DeepSeek’s card calls it the official release of DeepSeek-V4-Pro, superseding the preview version.Source: huggingface.co · checked 2026-11-19
  • The card describes the model as the quantized version of DeepSeek AI’s DeepSeek-V4-Pro-0813, made with NVIDIA’s Model Optimizer, and says it is ready for commercial or non-commercial use. It also says plainly that the underlying model is not owned or developed by NVIDIA and was built to a third party’s requirements, with a link to DeepSeek’s own model card.Source: huggingface.co · checked 2026-11-19
  • The card gives 1.65T total parameters and 49B activated for the Mixture-of-Experts network (class DeepseekV4ForCausalLM, with hybrid Compressed Sparse Attention and Heavily Compressed Attention), a maximum context of 1 million tokens, text in and text out, structured JSON output, tool calling, and three reasoning effort levels: low, high and max.Source: huggingface.co · checked 2026-11-19
  • What was quantized, per the card: only the weights and activations of the linear operators in the mixture-of-experts blocks, that is the routed experts. DeepSeek V4 already ships those experts in MXFP4, so NVIDIA carried the 4-bit weight values over unchanged and rewrote only the block scales; the checkpoint is therefore slightly larger than its source. DeepSeek’s DSpark speculative-decoding heads are carried through unquantized.Source: huggingface.co · checked 2026-11-19
  • Hugging Face’s API shows the file sizes of the NVIDIA repository add up to about 941 GB, and the total stored parameter count is 1,650,497,936,906.Source: huggingface.co · checked 2026-11-19
  • NVIDIA’s card lists NVIDIA Blackwell B200 as the supported hardware microarchitecture and as the test hardware, Linux as the preferred operating system, and SGLang and vLLM as the runtime engines. Its vLLM recipe uses tensor parallel size 8 with expert parallelism, a maximum model length of 400,000 tokens and an FP8 KV cache, and was validated on 8 B200 GPUs.Source: huggingface.co · checked 2026-11-19
  • The card lists the license as MIT and links DeepSeek’s LICENSE file; the repository’s own LICENSE file is the MIT text with "Copyright (c) 2023 DeepSeek". NVIDIA’s card does not add a separate NVIDIA license for this checkpoint.Source: huggingface.co · checked 2026-11-19
  • NVIDIA’s hosted trial page for the base model on build.nvidia.com has different wording: it says that service is governed by the NVIDIA API Trial Terms of Service, that use of the model is governed by the NVIDIA Open Model Agreement, and it gives MIT as additional information. That page covers NVIDIA’s hosted endpoint, not the downloadable NVFP4 checkpoint.Source: build.nvidia.com · checked 2026-11-19
  • The tool used to make it, NVIDIA Model Optimizer, is a separate open-source project that GitHub lists under the Apache-2.0 license, and the card points to a quantization recipe in that repository. The card also warns that the base model was trained on data containing toxic language and societal biases and can produce inaccurate or undesirable text.Source: github.com · checked 2026-11-19
  • OpenRouter’s list (checked November 19, 2026) has the unquantized base model as deepseek/deepseek-v4-pro-0813, "DeepSeek V4 Pro 0813", with a 1,048,576-token context and a list price of $0.111 per million input tokens and $4.44 per million output tokens. Its endpoints page shows 26 endpoints with input prices from $0.0999 to $1.98 per million and quantization labels fp4, mxfp4, fp8 or none; none of the labels says NVFP4. This NVFP4 checkpoint itself is not on the list.Source: openrouter.ai · checked 2026-11-19 · second-hand
  • The base model can be chatted with through OpenRouter, but this site does not list DeepSeek V4 Pro 0813 and cannot run this NVFP4 file, which is a downloadable weights file for people who run their own GPU servers. This site is a text chat with other models, such as DeepSeek V4.1 Flash and Kimi K3.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

DeepSeek V4 Pro 0813 NVFP4: what NVFP4 is, the 4-bit formats on NVIDIA Blackwell compared

  • NVIDIA’s technical blog (June 24, 2025) defines NVFP4 as a 4-bit floating-point format introduced with the NVIDIA Blackwell architecture: each value has 1 sign bit, 2 exponent bits and 1 mantissa bit (a range of about -6 to 6), and every block of 16 values shares one FP8 (E4M3) scale factor, with a second FP32 scale for the whole tensor.Source: developer.nvidia.com · checked 2026-11-19
  • Per the same blog, that costs about 4.5 bits per value, which cuts the memory footprint by roughly 3.5x against FP16 and roughly 1.8x against FP8. MXFP4, the format DeepSeek V4’s experts already use, shares one power-of-two scale across 32 values instead of an FP8 scale across 16. NVIDIA says its post-training quantization of DeepSeek-R1-0528 from FP8 to NVFP4 lost 1 percent or less on key language-modeling tasks; that is a claim about a different model.Source: developer.nvidia.com · checked 2026-11-19

Number formats

DeepSeek V4 Pro 0813 NVFP4: FP4, MXFP4 and NVFP4 as NVIDIA describes them

People who search for NVFP4 want to know whether it is just a smaller file or something that changes the answers. NVFP4 is a way of storing numbers in 4 bits plus a small shared scale for every 16 values, so the weights take about a quarter to a third of the memory of 16-bit weights. NVIDIA says the small blocks and the more precise scale keep the loss in accuracy low, and that Blackwell GPUs handle the format in hardware. The catch is the hardware: this is a format for Blackwell-generation NVIDIA GPUs, and the checkpoint’s card lists only the B200. If you do not own those, the practical way to use DeepSeek’s Pro model is a hosted one, or the cheaper DeepSeek V4.1 Flash on this site.

Our table: NVIDIA’s technical blog "Introducing NVFP4" (June 24, 2025) and the checkpoint’s Hugging Face card, read on November 19, 2026; the accuracy statements are NVIDIA’s and refer to its own tests · checked 2026-11-19
PropertyFP4 (E2M1)MXFP4NVFP4
Structure4 bits (1 sign, 2 exponent, 1 mantissa) plus a software scaling factor4 bits plus 1 shared power-of-two scale per 32-value block4 bits plus 1 shared FP8 scale per 16-value block
Scale precisionNot fixed by the formatE8M0 (power of two)E4M3 per block, plus an FP32 scale per tensor
Hardware-accelerated scalingNoYesYes
MemoryUp to 4x less than FP16Up to 4x less than FP16About 4.5 bits per value: roughly 3.5x less than FP16, 1.8x less than FP8
Accuracy risk (NVIDIA’s wording)Risk of a noticeable drop against FP8Risk of a noticeable drop against FP8Lower risk of a noticeable drop, particularly for larger models

Source: developer.nvidia.com

Our table

DeepSeek V4 Pro 0813 NVFP4 accuracy: NVIDIA’s own numbers against the MXFP4 original

  • NVIDIA’s card compares the NVFP4 checkpoint with the MXFP4 source on six benchmarks, all on SGLang at temperature 1.0, top-p 1.0 and max reasoning effort. GPQA Diamond 88.51 for the source against 88.42 for NVFP4, AA-LCR 68.67 against 69.33, τ²-Bench Telecom 96.49 against 98.25, SciCode 53.45 against 53.75, IFBench 76.53 against 75.68 and Terminal-Bench Hard 51.39 against 50.69. These are NVIDIA’s own figures, not tested by us.Source: huggingface.co · checked 2026-11-19
  • The card’s caveats on those numbers: GPQA Diamond, SciCode, τ²-Bench Telecom and Terminal-Bench Hard are single runs, IFBench is pooled over 4 runs per side and AA-LCR is a matched pair of 2 runs per side, and no error bars are printed. The card also says the vLLM configuration was validated on 8 B200 GPUs with vLLM 0.27.1 and 0.28.0 on a long-context run of about 115K-token prompts with no errors, but that the table itself was produced with SGLang, and that speculative decoding was not exercised in validation.Source: huggingface.co · checked 2026-11-19

Vendor benchmarks

DeepSeek V4 Pro 0813 NVFP4: NVFP4 checkpoint against the MXFP4 source checkpoint on six benchmarks

The question behind this table is whether 4-bit experts make the model worse. On NVIDIA’s numbers the answer is: not visibly. The NVFP4 scores sit within about two points of the source in both directions, higher on AA-LCR, τ²-Bench Telecom and SciCode and lower on GPQA Diamond, IFBench and Terminal-Bench Hard. Because most rows are single runs and the card prints no error bars, differences this small are inside normal run-to-run noise for these benchmarks, so read it as "no clear loss", not as "better". It also tests one setting, max reasoning effort on SGLang. Note that the base model compared here is already a mixed-precision 4-bit model, so this is MXFP4 against NVFP4, not against full 16-bit weights.

Our table: the evaluation table on the NVIDIA Hugging Face card, read on November 19, 2026; the scores are NVIDIA’s own and we did not run them, and the last column is our own subtraction, NVFP4 minus the source · checked 2026-11-19
BenchmarkMXFP4 (source)NVFP4Difference (our subtraction)
GPQA Diamond88.5188.42-0.09
AA-LCR68.6769.33+0.66
τ²-Bench Telecom96.4998.25+1.76
SciCode53.4553.75+0.30
IFBench76.5375.68-0.85
Terminal-Bench Hard51.3950.69-0.70

Source: huggingface.co

Our table

DeepSeek V4 Pro 0813 NVFP4 against the original DeepSeek-V4-Pro-0813

  • DeepSeek’s own card for DeepSeek-V4-Pro-0813 calls it the official release of V4-Pro, superseding the preview, with greatly enhanced agentic capabilities, built on the preview’s structure with a DSpark speculative-decoding module attached, and licensed under MIT. DeepSeek’s table, which it says used its own harness at max reasoning effort, lists Terminal Bench 2.1 at 87.9, DeepSWE at 62.7 and HLE at 42.7 without tools and 60.0 with tools.Source: huggingface.co · checked 2026-11-19
  • DeepSeek’s card recommends temperature 1.0 with top-p 0.95 for agentic use and 1.0 otherwise, and a maximum output of 384K tokens at the high and max effort levels. Its vLLM example serves the base model on a single node of 4 GB300 GPUs with DSpark speculative decoding switched on by one flag. Hugging Face’s API shows the base repository’s files add up to about 893 GB, against about 941 GB for the NVFP4 repository.Source: huggingface.co · checked 2026-11-19
  • DeepSeek’s API pricing page lists deepseek-v4-pro as DeepSeek-V4-Pro-0813 with a 1M context and up to 384K output, at $0.66 per million input tokens (cache miss) and $1.98 per million output tokens off-peak, and twice that at peak ($1.32 and $3.96). Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; cache-hit input is $0.022 off-peak and $0.044 at peak.Source: api-docs.deepseek.com · checked 2026-11-19

Side by side

DeepSeek V4 Pro 0813 NVFP4: NVIDIA’s NVFP4 checkpoint and DeepSeek’s base model: what differs and what does not

The NVFP4 checkpoint and the DeepSeek original are the same model with the same 1.65T parameters, 49B active and 1 million tokens of context; the difference is the format of the experts and who publishes the file. That is why this page is mostly about running a file, and why the base model’s facts (the benchmark numbers, the recommended sampling settings, the API price) come from DeepSeek, not from NVIDIA. If your goal is simply to ask DeepSeek V4 Pro questions, you do not need this checkpoint: the base model is on OpenRouter and in DeepSeek’s API. If your goal is to serve it yourself on Blackwell GPUs, the NVFP4 file is NVIDIA’s tested route. See also DeepSeek Harness, DeepSeek’s agent harness, which DeepSeek’s card names as its evaluation framework for the code-agent tasks.

Our table: the two Hugging Face cards and their API listings, DeepSeek’s API pricing page and OpenRouter’s list, read on November 19, 2026; file sizes are the sum of the files Hugging Face lists · checked 2026-11-19
ItemNVFP4 checkpoint (NVIDIA)DeepSeek-V4-Pro-0813 (DeepSeek)
PublishedAugust 27, 2026 on Hugging Face, by NVIDIARepository created August 13, 2026, by DeepSeek
Routed-expert formatNVFP4 (16-value blocks, E4M3 scale)MXFP4 (32-value blocks, E8M0 scale)
Parameters (total / active)1.65T / 49B1.65T / 49B (1,650,497,936,906 stored)
Files on Hugging FaceAbout 941 GBAbout 893 GB
LicenseMIT (text copyright DeepSeek)MIT
Hardware the card namesNVIDIA Blackwell B200, tested on 8 GPUs with SGLang and vLLMvLLM example on one node of 4 GB300 GPUs; SGLang example with tensor parallel 4
DSpark speculative decodingHeads kept unquantized, but not exercised in NVIDIA’s validationSwitched on with one flag in the vLLM and SGLang examples
Without your own GPUsNo hosted chat or API named on its cardDeepSeek’s API ($0.66 in, $1.98 out per million off-peak) and OpenRouter list it

Source: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.coSource: huggingface.coSource: api-docs.deepseek.comSource: openrouter.ai

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to DeepSeek V4 Pro 0813 NVFP4 you can use now

$0.30 / $1.2 (per million tokens)

  • DeepSeek released V4.1-Flash on September 10, 2026.Source: api-docs.deepseek.com · checked 2026-11-11
  • Only the Flash version of V4.1 exists: DeepSeek has published no V4.1 Pro, and the Pro model on its price page is still deepseek-v4-pro (V4-Pro-0813).Source: api-docs.deepseek.com · checked 2026-11-11
  • DeepSeek’s pricing page says the API model deepseek-flash is now served by the V4.1-Flash-1017 snapshot.Source: api-docs.deepseek.com · checked 2026-11-11

Kimi K3

Moonshot AI

$0.19 / $16 (per million tokens)

  • Moonshot’s Kimi Code changelog lists Kimi K3 as released on July 16, 2026, as open source. Moonshot’s blog said the full weights would follow by July 27, 2026.Source: kimi.com · checked 2026-11-16
  • Moonshot says K3 is designed for long-horizon coding, knowledge work and reasoning, and that it is available on Kimi.com, Kimi Work, Kimi Code and the Kimi API.Source: kimi.com · checked 2026-11-16
  • Mixture-of-experts model with 2.8T total parameters and 104B active per token, a 1,048,576-token context window, and image and video understanding. In the API, thinking is always on, with reasoning effort low, high or max (default max), and maximum output tokens default to 131,072 and can be raised to 1,048,576.Source: platform.kimi.ai · checked 2026-11-16

$0.08 / $3.6 (per million tokens)

  • A secondary report (DataNorth) says Z.ai released GLM-5.3 on August 14, 2026; Z.ai’s docs page calls it "Z.ai’s latest flagship model" but states no date.Source: datanorth.ai · checked 2026-11-15 · second-hand
  • Z.ai’s docs say GLM-5.3 advances complex software engineering and agent capabilities, with an emphasis on coding and cybersecurity performance.Source: docs.z.ai · checked 2026-11-15
  • Context window 1M tokens, maximum output 128K tokens; text-only input.Source: docs.z.ai · checked 2026-11-15

FAQ

DeepSeek V4 Pro 0813 NVFP4 FAQ

What is DeepSeek V4 Pro 0813 NVFP4?

It is a downloadable weights file, not a new model: NVIDIA took DeepSeek’s DeepSeek-V4-Pro-0813 and re-encoded its routed experts in the 4-bit NVFP4 format with its Model Optimizer tool. The model underneath is DeepSeek’s, with 1.65T total and 49B active parameters and a 1 million token context. NVIDIA published the file on Hugging Face on August 27, 2026.

What is NVFP4?

NVFP4 is a 4-bit floating-point format introduced with NVIDIA’s Blackwell GPUs. Each value is stored in 4 bits, and every block of 16 values shares an FP8 scale factor, with a second FP32 scale for the whole tensor. NVIDIA says this takes roughly 3.5x less memory than FP16 and 1.8x less than FP8, with lower accuracy risk than the older MXFP4 format.

What hardware does the NVFP4 checkpoint need?

NVIDIA’s card lists the NVIDIA Blackwell B200 as the supported and tested hardware, with SGLang and vLLM as engines, on Linux. The files add up to about 941 GB, and the card’s recipes use 8 GPUs (tensor parallel size 8). The card does not say it runs on older GPU generations, so assume a multi-GPU Blackwell server, not a desktop.

What license and terms apply to it?

NVIDIA’s card lists MIT and links DeepSeek’s MIT license, and the repository’s LICENSE file is the MIT text with DeepSeek as copyright holder; the card says the model is ready for commercial or non-commercial use. NVIDIA’s separate hosted trial page for the base model refers to the NVIDIA Open Model Agreement and trial terms, but that covers its hosted service, not this download.

How accurate is the NVFP4 version compared with the original?

By NVIDIA’s own table, which we did not run, the NVFP4 scores stay within about two points of the MXFP4 original on six benchmarks: GPQA Diamond 88.42 against 88.51, AA-LCR 69.33 against 68.67, and Terminal-Bench Hard 50.69 against 51.39. Most rows are single runs with no error bars, so it shows no clear loss, not a gain.

How do I run it?

NVIDIA’s card gives two recipes. For SGLang it launches nvidia/DeepSeek-V4-Pro-0813-NVFP4 with --tp 8, --trust-remote-code and the deepseekv4 parsers. For vLLM it uses tensor parallel size 8, expert parallelism, a 400,000-token maximum length and an FP8 KV cache, validated on 8 B200 GPUs. Speculative decoding was not tested for this checkpoint, so enable it only after your own checks.

Can I chat with DeepSeek V4 Pro here, and what can I use instead?

Not this file, and not the Pro model on this site. The unquantized base model, DeepSeek V4 Pro 0813, is on OpenRouter and in DeepSeek’s API, which lists $0.66 per million input and $1.98 per million output tokens off-peak. On this site you can chat with DeepSeek V4.1 Flash: a typical reply costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Kimi K3 is another large model.

Is this the official DeepSeek V4 Pro 0813 NVFP4 website?

No. This page is unofficial and not affiliated with or endorsed by NVIDIA or DeepSeek. Facts link to the pages we read.