NVFP4 models for the NVIDIA RTX PRO 6000 Blackwell: checkpoint list for October 2026

Which NVFP4 checkpoints exist for the RTX PRO 6000

As of 24 October 2026, NVIDIA's Hugging Face account holds 85 repositories with NVFP4 in the name, and 22 of them are current language models a company would serve on an NVIDIA RTX PRO 6000 Blackwell Server Edition. Their checkpoints range from 6.0 GB for Llama 3.1 8B to 595.2 GB for Kimi K2.6; thirteen fit one 96 GB card with room for a KV cache, the largest need a server with eight cards. Nearly every model card names NVIDIA Blackwell as the target, and none names the RTX PRO 6000 as the hardware it was tested on.

NVFP4 is the 4-bit format that Blackwell Tensor Cores compute natively, and the RTX PRO 6000 in all three editions is a Blackwell card with compute capability 12.0. How the format works is explained in our guide to FP8, NVFP4, MXFP4 and INT4; which engine releases run it on these cards is in our guide to SM120 software support. This page is the list that both refer to: which ready-made NVFP4 checkpoints exist, how large they are and how many cards each one needs. It is a snapshot of one day and will age quickly: the most recently updated repositories in the search were changed on 23 October 2026.

How the list was built

The figures come from three places, all read on 24 October 2026.

• Size. The sum of the safetensors files in each repository's file list, in decimal GB and in GiB. Tokeniser and configuration files are left out; they add a few megabytes.
• Card count. Our users-per-card rule: 90 per cent of the 95.6 GiB the driver reports, less 3 GiB per card for activations and the runtime. That leaves 83.0 GiB per card for weights and cache, 166.1 GiB on two cards, 332.2 GiB on four and 664.3 GiB on eight. The list gives the smallest of these that holds the weights and what remains for the KV cache.
• Card details. Test hardware, supported engines, what is quantised and the licence come from each NVIDIA model card.

The search also returns variants with suffixes such as DSpark, DFlash, QAD or MLPerf, older generations of the same families and embedding, image and video models; they are not in the tables. The checkpoint size counts everything in the files, including vision encoders and layers kept in 8 or 16 bits, which is why it is larger than parameters × 0.5625 bytes.

NVFP4 checkpoints that fit one RTX PRO 6000

Hugging Face file lists (sum of safetensors files) and model cards of the nvidia/ repositories named, read 24 October 2026; free memory by our users-per-card rule (83.0 GiB per card less the weights). Some names shortened: the Nemotron repositories start with NVIDIA- and end in -NVFP4, Llama 4 Scout is Llama-4-Scout-17B-16E-Instruct-NVFP4, the 49B model Llama-3_3-Nemotron-Super-49B-v1_5-NVFP4. For the two Super models the card gives a GPU list rather than a test-hardware field.

The last three rows load on one card, but the space left for the cache is small. How many conversations that holds depends on the model's attention design: the Nemotron cards describe Mamba layers, which keep no per-token KV cache, so their cache per token is smaller than that of a dense transformer, and a fixed figure per session would mislead. Read the cache size vLLM prints at start-up, as our guide to vLLM KV cache size and maximum concurrency describes. For a service with more than a handful of parallel users, plan two cards for these three.

For the dense 70B model the arithmetic is published in our guide to users per RTX PRO 6000: with NVFP4 weights and an FP8 KV cache, one card holds 34 full 8K sessions of Llama 3.3 70B.

NVFP4 checkpoints for two, four or eight cards

as above, names shortened in the same way; usable memory by our rule 166.1 GiB on two cards, 332.2 GiB on four, 664.3 GiB on eight. DeepSeek-V4-Flash-0731 leaves 2.6 GiB on two cards and Nemotron 3 Ultra 4.1 GiB on four, so the table gives the next size up. The Nemotron 3 Ultra card lists H100, H200, GB200, GB300, B200 and B300.

Two models sit on a boundary. DeepSeek-V4-Flash-0731 in NVIDIA's NVFP4 version just fits two cards by weight and leaves no cache, so it is a four-card model on the RTX PRO 6000. Nemotron 3 Ultra is the same case one step up: 328.1 GiB against 332.2 GiB on four cards. Both are better planned on the next server size, as our guides to the four-card and eight-card RTX PRO 6000 Server Edition server do for their models.

The cards' own deployment examples assume far more memory per GPU. NVIDIA serves Qwen3.8-Flash-Next and DeepSeek-V4-Flash-0731 with eight GPUs in its sample commands, GLM-5.3-Flash and Kimi K2.6 with four, and the Nemotron 3 Ultra card states a minimum of four B200 or eight H100. On RTX PRO 6000 cards the layouts follow from 96 GB per card and PCIe without NVLink: tensor parallelism within a node, sized by the tables above.

We build servers with two, four or eight RTX PRO 6000 Server Edition cards to order; send us the checkpoint you want to run and your number of users, and our team replies with the card count and memory configuration it needs.

What the model cards say about hardware and engines

All cards in the list but one name NVIDIA Blackwell under supported or compatible microarchitecture; the Llama-3.3-Nemotron-Super-49B NVFP4 card lists only Ampere and Hopper as recommended. Several also name Hopper or Ampere: the Llama 4 Scout card lists Blackwell, Hopper and Ampere, Qwen3.6-35B-A3B lists Hopper and Blackwell, and the Nemotron 3.5 Lightning card lists "NVIDIA Hopper (NVFP4 / W4A16); NVIDIA Ampere (W4A16)". On Hopper and Ampere these checkpoints run through weight-only kernels, as our guide to FP4 models on the H200 NVL and Ada GPUs explains. The Qwen3.8-Flash-Next card is the narrowest: it names "NVIDIA Blackwell B200" and "NVIDIA Blackwell B300" as supported hardware.

The test hardware is where the RTX PRO 6000 is missing. The B200 and B300 are compute capability 10.0 and 10.3, the RTX PRO 6000 is 12.0, and engines ship separate kernels for the two. A checkpoint tested on a B200 has shown that the format and the recipe work, not that every kernel the engine picks on SM120 is in place. For the large mixture-of-experts models, check the engine's release notes for the model family on SM120 before you order, as our SM120 guide records for DeepSeek-class models.

The engines named differ by card. The older checkpoints (Llama 3.3 70B, Qwen3-32B, Qwen3-235B-A22B-2507, Llama 4 Scout) name TensorRT-LLM; the newer ones name vLLM, SGLang or both. Model Optimizer's documentation states that its exported checkpoints deploy on TensorRT-LLM, vLLM and SGLang, and in vLLM they load with quantization="modelopt".

What is quantised, and why sizes vary

"NVFP4" in a repository name covers several recipes. The cards state them:

the NVIDIA model cards named, read 24 October 2026.

Recipes that keep attention or shared experts in 8 or 16 bits trade memory for accuracy, and the sizes show it. Mistral Medium 3.5 has 128B parameters and an NVFP4 checkpoint of 95.2 GB, about 0.74 bytes per parameter, because the MLP of five of its 88 decoder layers (0 to 87), the attention projections and the KV cache stay in FP8. DeepSeek-V4-Flash-0731 shows the extreme: DeepSeek already publishes it with FP4 experts and FP8 elsewhere at 167 GB, as our DeepSeek hardware guide lists, and NVIDIA's NVFP4 version is 175.6 GB. It exists to give the engines' NVFP4 kernels a checkpoint in their format and saves no memory; on two cards only DeepSeek's own release leaves room for a cache (166.1 GiB against about 155.5 GiB of weights, by our estimate).

Several cards also publish accuracy against the BF16 model, measured by NVIDIA on its own hardware:

NVIDIA model cards of Gemma-4-31B-IT-NVFP4, Qwen3.8-27B-NVFP4, Qwen3.6-35B-A3B-NVFP4 and NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, read 24 October 2026. Lightning = Nemotron 3.5 Lightning.

Most differences are below one point; the Lightning card shows that coding and long-context tasks can lose more, nearly three points on SWE-bench Multilingual. Run your own evaluation set on the checkpoint before it replaces an FP8 or BF16 model, as our guide to LLM evaluation before model upgrades describes.

Licences to read before deployment

NVIDIA's NVFP4 repositories carry two layers of terms: NVIDIA's own and those of the original model. The Apache 2.0 and MIT models in the list (Gemma 4, Qwen3, Qwen3.5, Qwen3.6, Qwen3.8-27B, DeepSeek-V4-Flash, GLM-5.3-Flash) carry no use restrictions beyond those licences. Several others do:

• Llama 4 Scout. The card names the NVIDIA Open Model License and the Llama 4 Community License. Meta's Llama 4 Acceptable Use Policy states: "With respect to any multimodal models included in Llama 4, the rights granted under Section 1(a) of the Llama 4 Community License Agreement are not being granted to you if you are an individual domiciled in, or a company with a principal place of business in, the European Union." It exempts end users of a product that incorporates such a model. Llama 4 Scout accepts text and images, per Meta's model card. A company based in the EU should not plan on this checkpoint without its legal counsel's assessment.
• Mistral Medium 3.5. The card names Mistral's Modified MIT license. As quoted in our Mistral hardware guide, its second section grants no rights to a company whose global consolidated monthly revenue exceeded USD 20 million in the preceding month.
• Qwen3.8-Flash-Next. NVIDIA Open Model License plus the Qwen Community License 1.0, which, as our guide to open LLM licences sets out, requires a separate licence for Model as a Service and AI work assistant businesses.
• Kimi K2.6 and GLM-5.3. Kimi K2.6 adds Moonshot AI's Modified MIT License, with a display requirement for very large commercial services; GLM-5.3 is governed by the NVIDIA Open Model Agreement, which, as the card states, does not modify or supersede the GLM-5.3 License or its conditions.
• Llama 3.1 and 3.3. Llama 3.1 8B and Llama 3.3 70B name the NVIDIA Open Model License with Meta's Llama 3 or Llama 3.3 licence as additional terms; the conditions of Meta's licences, among them the clause for services above 700 million monthly users, are in our licence guide.
• Nemotron. Nemotron 3 Super uses the NVIDIA Nemotron Open Model License; Nemotron 3 Ultra, 3.5 Super VL and 3.5 Lightning the OpenMDW License Agreement 1.1.

This is general information from the licence texts, not legal advice; your legal counsel confirms the terms for your use.

How to choose a checkpoint from the list

1. Choose the model first. Pick it by task and licence; then check whether NVIDIA's NVFP4 version exists in the tables above. Memory needs in other formats are in our LLM hardware requirements by model.
2. Size by the table. The card count above is the minimum for weights plus some cache. Add the users and context length you need, by the rule in our users guide.
3. Read the recipe. Model Optimizer's documentation recommends experts-only and MLP-only recipes for higher NVFP4 accuracy; they make larger checkpoints than recipes that quantise every linear layer.
4. Check the engine on SM120. Use the engine the card names, at a release that lists SM120 support for that model family, and read the start-up log for fallback warnings.
5. Evaluate. Compare the NVFP4 checkpoint with the FP8 or BF16 model on your own prompts. If no NVFP4 checkpoint exists for your model, our guide to quantising an LLM to NVFP4 or FP8 covers producing one.

Our team sizes the server for the checkpoint you pick and quotes it in writing; tell us which models from the list you plan to serve, and we reply with a configuration for two, four or eight cards.

FAQ