Skip to content
Local AI Wizard
Your machine
Theme

NVLM-D-72B

NVIDIA · nvidia/NVLM-D-72B

Weights license: Publisher terms: CC BY-NC 4.0 · code and components may use other terms

Reported by the publisher12 resultsNVIDIA's model card · read 2026-10-25 · not ranked · see them
Memory needed~60 GBQ4 weights + working room, before the machine's reserve
Parameters79.4 BHub safetensors total
Repository updated2025-01-14A commit is not a release date

Quality

No comparable score

We track no public, comparable benchmark that covers this model yet. No score is inferred.

Reported by the publisher

12 results · NVIDIA's model card · read 2026-10-25
MathVista
65.2Vision-language Benchmarks
OCRBench
852Vision-language Benchmarks
AI2D
94.2Vision-language Benchmarks
ChartQA
86.0Vision-language Benchmarks
DocVQA
92.6Vision-language Benchmarks
TextVQA
82.6Vision-language Benchmarks
RealWorldQA
69.5Vision-language Benchmarks
VQAv2
85.4Vision-language Benchmarks
Every reported result12
MMLU
81.7Text-only Benchmarks
GSM8K
93.2Text-only Benchmarks
MATH
73.1Text-only Benchmarks
HumanEval
89.0Text-only Benchmarks

Values as printed on the card, with the publisher's own settings (card section: Model Details › Benchmark Results › Vision-language Benchmarks; Model Details › Benchmark Results › Text-only Benchmarks); column “NVLM-D 1.0 72B (Huggingface)”. Not comparable across publishers and not used in any ranking. Model card

Runs on

Fast memory holds it 15

The planning estimate fits the memory the model can use quickly: graphics card VRAM, or unified memory within the macOS GPU share.

Show all 15

Within memory, outside the fast part 1

Within the machine's memory, outside its fast part: on a graphics card the layers that do not fit run on the CPU from PC RAM (the listed RAM, or 32 GB assumed); on a Mac, above the macOS default GPU share. Not a speed figure.

One memory tier short 7

The same machines with the next memory size up would hold it in fast memory.

Show all 7

Hardware profiles grouped by the Q4 planning estimate on each machine (dedicated GPUs reserve 2 GB; unified and system memory 25%, 2 to 8 GB): the fast memory holds it, it is within memory but outside the fast part (layers offloaded to PC RAM, or above the macOS GPU share), or the machine is one memory tier short. Not a measured fit or speed.

Consider it for

NVIDIA released NVLM-D-72B on 2024-09-17: a vision-language model that answers about images, with 79.4B parameters on the Hugging Face count and the Publisher terms: CC BY-NC 4.0 license. Languages the publisher declares: en.

Check first

License: commercial use not allowed. At Q4 the weights load about 59.5 GB before each machine's reserve; image resolution, video frames and context add memory, and the vision encoder needs support in the runtime you pick.

Measured speed

No published measurement for this checkpoint

Tokens per second are shown only when a source publishes them with hardware, build and setup. They are never estimated.

How to run

Runtime compatibility and tokens per second have not been measured for this entry. Published metadata and a memory estimate do not prove a successful local run.

Library in source: transformers

Documented ways to run it

Copied as printed in the linked source (read 2026-10-18); not run by this site.

Transformers · model card · Inference
import torch
from transformers import AutoTokenizer, AutoModel
import math
from PIL import Image
import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode


def split_model():
    device_map = {}
    world_size = torch.cuda.device_count()
    num_layers = 80
    # Since the first GPU will be used for ViT, treat it as half a GPU.
    num_layers_per_gpu = math.ceil(num_layers / (world_size - 0.5))
    num_layers_per_gpu = [num_layers_per_gpu] * world_size
    num_layers_per_gpu[0] = math.ceil(num_layers_per_gpu[0] * 0.5)
    layer_cnt = 0
    for i, num_layer in enumerate(num_layers_per_gpu):
        for j in range(num_layer):
            device_map[f'language_model.model.layers.{layer_cnt}'] = i
            layer_cnt += 1
    device_map['vision_model'] = 0
    device_map['mlp1'] = 0
    device_map['language_model.model.tok_embeddings'] = 0
    device_map['language_model.model.embed_tokens'] = 0
    device_map['language_model.output'] = 0
    device_map['language_model.model.norm'] = 0
    device_map['language_model.lm_head'] = 0
    device_map['language_model.model.rotary_emb'] = 0
    device_map[f'language_model.model.layers.{num_layers - 1}'] = 0

    return device_map


IMAGENET_MEAN = (0.485, 0.456, 0.406)
IMAGENET_STD = (0.229, 0.224, 0.225)


def build_transform(input_size):
    MEAN, STD = IMAGENET_MEAN, IMAGENET_STD
    transform = T.Compose([
        T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
        T.Resize((input_size, input_size), interpolation=InterpolationMode.BICUBIC),
        T.ToTensor(),
        T.Normalize(mean=MEAN, std=STD)
    ])
    return transform


def find_closest_aspect_ratio(aspect_ratio, target_ratios, width, height, image_size):
    best_ratio_diff = float('inf')
    best_ratio = (1, 1)
    area = width * height
    for ratio in target_ratios:
        target_aspect_ratio = ratio[0] / ratio[1]
        ratio_diff = abs(aspect_ratio - target_aspect_ratio)
        if ratio_diff < best_ratio_diff:
            best_ratio_diff = ratio_diff
            best_ratio = ratio
        elif ratio_diff == best_ratio_diff:
            if area > 0.5 * image_size * image_size * ratio[0] * ratio[1]:
                best_ratio = ratio
    return best_ratio


def dynamic_preprocess(image, min_num=1, max_num=12, image_size=448, use_thumbnail=False):
    orig_width, orig_height = image.size
    aspect_ratio = orig_width / orig_height

    # calculate the existing image aspect ratio
    target_ratios = set(
        (i, j) for n in range(min_num, max_num + 1) for i in range(1, n + 1) for j in range(1, n + 1) if
        i * j <= max_num and i * j >= min_num)
    target_ratios = sorted(target_ratios, key=lambda x: x[0] * x[1])

    # find the closest aspect ratio to the target
    target_aspect_ratio = find_closest_aspect_ratio(
        aspect_ratio, target_ratios, orig_width, orig_height, image_size)

    # calculate the target width and height
    target_width = image_size * target_aspect_ratio[0]
    target_height = image_size * target_aspect_ratio[1]
    blocks = target_aspect_ratio[0] * target_aspect_ratio[1]

    # resize the image
    resized_img = image.resize((target_width, target_height))
    processed_images = []
    for i in range(blocks):
        box = (
            (i % (target_width // image_size)) * image_size,
            (i // (target_width // image_size)) * image_size,
            ((i % (target_width // image_size)) + 1) * image_size,
            ((i // (target_width // image_size)) + 1) * image_size
        )
        # split the image
        split_img = resized_img.crop(box)
        processed_images.append(split_img)
    assert len(processed_images) == blocks
    if use_thumbnail and len(processed_images) != 1:
        thumbnail_img = image.resize((image_size, image_size))
        processed_images.append(thumbnail_img)
    return processed_images


def load_image(image_file, input_size=448, max_num=12):
    image = Image.open(image_file).convert('RGB')
    transform = build_transform(input_size=input_size)
    images = dynamic_preprocess(image, image_size=input_size, use_thumbnail=True, max_num=max_num)
    pixel_values = [transform(image) for image in images]
    pixel_values = torch.stack(pixel_values)
    return pixel_values

path = "nvidia/NVLM-D-72B"
device_map = split_model()
model = AutoModel.from_pretrained(
    path,
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    use_flash_attn=False,
    trust_remote_code=True,
    device_map=device_map).eval()

print(model)

tokenizer = AutoTokenizer.from_pretrained(path, trust_remote_code=True, use_fast=False)
generation_config = dict(max_new_tokens=1024, do_sample=False)

# pure-text conversation
question = 'Hello, who are you?'
response, history = model.chat(tokenizer, None, question, generation_config, history=None, return_history=True)
print(f'User: {question}\nAssistant: {response}')

# single-image single-round conversation
pixel_values = load_image('path/to/your/example/image.jpg', max_num=6).to(
    torch.bfloat16)
question = '<image>\nPlease describe the image shortly.'
response = model.chat(tokenizer, pixel_values, question, generation_config)
print(f'User: {question}\nAssistant: {response}')

Choose my computer

Technical details

What the source and the catalog store for this checkpoint. Missing fields are left out, never zero.

Checkpoint / source
nvidia/NVLM-D-72B
Task
image-text-to-text
Architecture
NVLM_D
Library in source
transformers
Weight license
Publisher terms: CC BY-NC 4.0nvidia/NVLM-D-72B model card (Hugging Face) · read 2026-10-18source
Commercial use
Not allowed CC BY-NC 4.0 text (SPDX) · read 2026-10-18license text
Languages
enModel card metadata · read 2026-10-25source
Released
NVLM-D-72B model card, 'Today (September 17th, 2024), we introduce NVLM 1.0 ... open-sourcing NVLM-1.0-D-72B' (Hugging Face) · read 2026-10-18source
Every stored field12 fields
Revision checked
11e9e0f8d265d65ddaa8f9445b6f55fdec9d686f
Architecture
NVLM_D
Library
transformers
Hub task tag
image-text-to-text
Formats
transformers, safetensors, pytorch
Access
Open download
Publisher kind
Official
Hub parameter count
79,379,593,344safetensors.total
Repository created
2024-09-30creation is not the model's release
Repository updated
2025-01-14a commit is not a release date
Hub downloads (30 days) · likes
8,951 · 780this repository only · read 2026-10-25
Catalog state
Hub config or weights found; no inventory benchmark link

Sources

Identity and published specifications come from the checkpoint page. Benchmark scores belong to the named release and do not measure speed on your computer.

Curated facts checked: 2026-10-18. The license and release date cite the publisher's sources; parameters, files and task come from the Hugging Face record, and memory is a labelled estimate from its parameter count. A value no source states stays empty; a repository commit is not a model release date.