kosmos-2-patch14-224 by Microsoft: Open-Weight Model

Runs On

What it takes to serve kosmos-2-patch14-224 (1.7B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

Precision	Weights	Memory needed	Cheapest setup	Per hour	Also fits
16-bit	3.3 GB	4.0 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
8-bit	1.7 GB	2.0 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00
4-bit	0.8 GB	1.0 GB	1x MI300X (192 GB) Vultr	$1.85	1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Sep 18, 2026.

Model Card

By Microsoft, published under mit, revision e91cfbcb4ce0.

This Hub repository contains a HuggingFace's transformers implementation of the original Kosmos-2 model from Microsoft. Use the code below to get started with the model. This model is capable of performing different tasks through changing the prompts. First, let's define a function to run a prompt. Here are the tasks Kosmos-2 could perform: Once you have the entities, you can use the following helper function to draw their bounding bboxes on the image

Read Microsoft's full model card

Kosmos-2: Grounding Multimodal Large Language Models to the World

[An image of a snowman warming himself by a fire.]

This Hub repository contains a HuggingFace's transformers implementation of the original Kosmos-2 model from Microsoft.

How to Get Started with the Model

Use the code below to get started with the model.

import requests

from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq


model = AutoModelForVision2Seq.from_pretrained("microsoft/kosmos-2-patch14-224")
processor = AutoProcessor.from_pretrained("microsoft/kosmos-2-patch14-224")

prompt = "<grounding>An image of"

url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.png"
image = Image.open(requests.get(url, stream=True).raw)

# The original Kosmos-2 demo saves the image first then reload it. For some images, this will give slightly different image input and change the generation outputs.
image.save("new_image.jpg")
image = Image.open("new_image.jpg")

inputs = processor(text=prompt, images=image, return_tensors="pt")

generated_ids = model.generate(
    pixel_values=inputs["pixel_values"],
    input_ids=inputs["input_ids"],
    attention_mask=inputs["attention_mask"],
    image_embeds=None,
    image_embeds_position_mask=inputs["image_embeds_position_mask"],
    use_cache=True,
    max_new_tokens=128,
)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

# Specify `cleanup_and_extract=False` in order to see the raw model generation.
processed_text = processor.post_process_generation(generated_text, cleanup_and_extract=False)

print(processed_text)
# `<grounding> An image of<phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> warming himself by<phrase> a fire</phrase><object><patch_index_0005><patch_index_0911></object>.`

# By default, the generated  text is cleanup and the entities are extracted.
processed_text, entities = processor.post_process_generation(generated_text)

print(processed_text)
# `An image of a snowman warming himself by a fire.`

print(entities)
# `[('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a fire', (41, 47), [(0.171875, 0.015625, 0.484375, 0.890625)])]`

Tasks

This model is capable of performing different tasks through changing the prompts.

First, let's define a function to run a prompt.

Click to expand

import requests

from PIL import Image
from transformers import AutoProcessor, AutoModelForVision2Seq


model = AutoModelForVision2Seq.from_pretrained("microsoft/kosmos-2-patch14-224")
processor = AutoProcessor.from_pretrained("microsoft/kosmos-2-patch14-224")

url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.png"
image = Image.open(requests.get(url, stream=True).raw)

def run_example(prompt):

    inputs = processor(text=prompt, images=image, return_tensors="pt")
    generated_ids = model.generate(
      pixel_values=inputs["pixel_values"],
      input_ids=inputs["input_ids"],
      attention_mask=inputs["attention_mask"],
      image_embeds=None,
      image_embeds_position_mask=inputs["image_embeds_position_mask"],
      use_cache=True,
      max_new_tokens=128,
    )
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    _processed_text = processor.post_process_generation(generated_text, cleanup_and_extract=False)
    processed_text, entities = processor.post_process_generation(generated_text)

    print(processed_text)
    print(entities)
    print(_processed_text)

Here are the tasks Kosmos-2 could perform:

Click to expand

### Multimodal Grounding #### • Phrase Grounding

prompt = "<grounding><phrase> a snowman</phrase>"
run_example(prompt)

# a snowman is warming himself by the fire
# [('a snowman', (0, 9), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('the fire', (32, 40), [(0.203125, 0.015625, 0.453125, 0.859375)])]

# <grounding><phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> is warming himself by<phrase> the fire</phrase><object><patch_index_0006><patch_index_0878></object>

#### • Referring Expression Comprehension

prompt = "<grounding><phrase> a snowman next to a fire</phrase>"
run_example(prompt)

# a snowman next to a fire
# [('a snowman next to a fire', (0, 24), [(0.390625, 0.046875, 0.984375, 0.828125)])]

# <grounding><phrase> a snowman next to a fire</phrase><object><patch_index_0044><patch_index_0863></object>

### Multimodal Referring #### • Referring expression generation

prompt = "<grounding><phrase> It</phrase><object><patch_index_0044><patch_index_0863></object> is"
run_example(prompt)

# It is snowman in a hat and scarf
# [('It', (0, 2), [(0.390625, 0.046875, 0.984375, 0.828125)])]

# <grounding><phrase> It</phrase><object><patch_index_0044><patch_index_0863></object> is snowman in a hat and scarf

### Perception-Language Tasks #### • Grounded VQA

prompt = "<grounding> Question: What is special about this image? Answer:"
run_example(prompt)

# Question: What is special about this image? Answer: The image features a snowman sitting by a campfire in the snow.
# [('a snowman', (71, 80), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a campfire', (92, 102), [(0.109375, 0.640625, 0.546875, 0.984375)])]

# <grounding> Question: What is special about this image? Answer: The image features<phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> sitting by<phrase> a campfire</phrase><object><patch_index_0643><patch_index_1009></object> in the snow.

#### • Grounded VQA with multimodal referring via bounding boxes

prompt = "<grounding> Question: Where is<phrase> the fire</phrase><object><patch_index_0005><patch_index_0911></object> next to? Answer:"
run_example(prompt)

# Question: Where is the fire next to? Answer: Near the snowman.
# [('the fire', (19, 27), [(0.171875, 0.015625, 0.484375, 0.890625)]), ('the snowman', (50, 61), [(0.390625, 0.046875, 0.984375, 0.828125)])]

# <grounding> Question: Where is<phrase> the fire</phrase><object><patch_index_0005><patch_index_0911></object> next to? Answer: Near<phrase> the snowman</phrase><object><patch_index_0044><patch_index_0863></object>.

### Grounded Image captioning #### • Brief

prompt = "<grounding> An image of"
run_example(prompt)

# An image of a snowman warming himself by a campfire.
# [('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a campfire', (41, 51), [(0.109375, 0.640625, 0.546875, 0.984375)])]

# <grounding> An image of<phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> warming himself by<phrase> a campfire</phrase><object><patch_index_0643><patch_index_1009></object>.

#### • Detailed

prompt = "<grounding> Describe this image in detail:"
run_example(prompt)

# Describe this image in detail: The image features a snowman sitting by a campfire in the snow. He is wearing a hat, scarf, and gloves, with a pot nearby and a cup nearby. The snowman appears to be enjoying the warmth of the fire, and it appears to have a warm and cozy atmosphere.
# [('a campfire', (71, 81), [(0.171875, 0.015625, 0.484375, 0.984375)]), ('a hat', (109, 114), [(0.515625, 0.046875, 0.828125, 0.234375)]), ('scarf', (116, 121), [(0.515625, 0.234375, 0.890625, 0.578125)]), ('gloves', (127, 133), [(0.515625, 0.390625, 0.640625, 0.515625)]), ('a pot', (140, 145), [(0.078125, 0.609375, 0.265625, 0.859375)]), ('a cup', (157, 162), [(0.890625, 0.765625, 0.984375, 0.984375)])]

# <grounding> Describe this image in detail: The image features a snowman sitting by<phrase> a campfire</phrase><object><patch_index_0005><patch_index_1007></object> in the snow. He is wearing<phrase> a hat</phrase><object><patch_index_0048><patch_index_0250></object>,<phrase> scarf</phrase><object><patch_index_0240><patch_index_0604></object>, and<phrase> gloves</phrase><object><patch_index_0400><patch_index_0532></object>, with<phrase> a pot</phrase><object><patch_index_0610><patch_index_0872></object> nearby and<phrase> a cup</phrase><object><patch_index_0796><patch_index_1023></object> nearby. The snowman appears to be enjoying the warmth of the fire, and it appears to have a warm and cozy atmosphere.

Draw the bounding bboxes of the entities on the image

Once you have the entities, you can use the following helper function to draw their bounding bboxes on the image:

Click to expand

import cv2
import numpy as np
import os
import requests
import torch
import torchvision.transforms as T

from PIL import Image


def is_overlapping(rect1, rect2):
    x1, y1, x2, y2 = rect1
    x3, y3, x4, y4 = rect2
    return not (x2 < x3 or x1 > x4 or y2 < y3 or y1 > y4)


def draw_entity_boxes_on_image(image, entities, show=False, save_path=None):
    """_summary_
    Args:
        image (_type_): image or image path
        collect_entity_location (_type_): _description_
    """
    if isinstance(image, Image.Image):
        image_h = image.height
        image_w = image.width
        image = np.array(image)[:, :, [2, 1, 0]]
    elif isinstance(image, str):
        if os.path.exists(image):
            pil_img = Image.open(image).convert("RGB")
            image = np.array(pil_img)[:, :, [2, 1, 0]]
            image_h = pil_img.height
            image_w = pil_img.width
        else:
            raise ValueError(f"invaild image path, {image}")
    elif isinstance(image, torch.Tensor):
        image_tensor = image.cpu()
        reverse_norm_mean = torch.tensor([0.48145466, 0.4578275, 0.40821073])[:, None, None]
        reverse_norm_std = torch.tensor([0.26862954, 0.26130258, 0.27577711])[:, None, None]
        image_tensor = image_tensor * reverse_norm_std + reverse_norm_mean
        pil_img = T.ToPILImage()(image_tensor)
        image_h = pil_img.height
        image_w = pil_img.width
        image = np.array(pil_img)[:, :, [2, 1, 0]]
    else:
        raise ValueError(f"invaild image format, {type(image)} for {image}")

    if len(entities) == 0:
        return image

    new_image = image.copy()
    previous_bboxes = []
    # size of text
    text_size = 1
    # thickness of text
    text_line = 1  # int(max(1 * min(image_h, image_w) / 512, 1))
    box_line = 3
    (c_width, text_height), _ = cv2.getTextSize("F", cv2.FONT_HERSHEY_COMPLEX, text_size, text_line)
    base_height = int(text_height * 0.675)
    text_offset_original = text_height - base_height
    text_spaces = 3

    for entity_name, (start, end), bboxes in entities:
        for (x1_norm, y1_norm, x2_norm, y2_norm) in bboxes:
            orig_x1, orig_y1, orig_x2, orig_y2 = int(x1_norm * image_w), int(y1_norm * image_h), int(x2_norm * image_w), int(y2_norm * image_h)
            # draw bbox
            # random color
            color = tuple(np.random.randint(0, 255, size=3).tolist())
            new_image = cv2.rectangle(new_image, (orig_x1, orig_y1), (orig_x2, orig_y2), color, box_line)

            l_o, r_o = box_line // 2 + box_line % 2, box_line // 2 + box_line % 2 + 1

            x1 = orig_x1 - l_o
            y1 = orig_y1 - l_o

            if y1 < text_height + text_offset_original + 2 * text_spaces:
                y1 = orig_y1 + r_o + text_height + text_offset_original + 2 * text_spaces
                x1 = orig_x1 + r_o

            # add text background
            (text_width, text_height), _ = cv2.getTextSize(f"  {entity_name}", cv2.FONT_HERSHEY_COMPLEX, text_size, text_line)
            text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2 = x1, y1 - (text_height + text_offset_original + 2 * text_spaces), x1 + text_width, y1

            for prev_bbox in previous_bboxes:
                while is_overlapping((text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2), prev_bbox):
                    text_bg_y1 += (text_height + text_offset_original + 2 * text_spaces)
                    text_bg_y2 += (text_height + text_offset_original + 2 * text_spaces)
                    y1 += (text_height + text_offset_original + 2 * text_spaces)

                    if text_bg_y2 >= image_h:
                        text_bg_y1 = max(0, image_h - (text_height + text_offset_original + 2 * text_spaces))
                        text_bg_y2 = image_h
                        y1 = image_h
                        break

            alpha = 0.5
            for i in range(text_bg_y1, text_bg_y2):
                for j in range(text_bg_x1, text_bg_x2):
                    if i < image_h and j < image_w:
                        if j < text_bg_x1 + 1.35 * c_width:
                            # original color
                            bg_color = color
                        else:
                            # white
                            bg_color = [255, 255, 255]
                        new_image[i, j] = (alpha * new_image[i, j] + (1 - alpha) * np.array(bg_color)).astype(np.uint8)

            cv2.putText(
                new_image, f"  {entity_name}", (x1, y1 - text_offset_original - 1 * text_spaces), cv2.FONT_HERSHEY_COMPLEX, text_size, (0, 0, 0), text_line, cv2.LINE_AA
            )
            # previous_locations.append((x1, y1))
            previous_bboxes.append((text_bg_x1, text_bg_y1, text_bg_x2, text_bg_y2))

    pil_image = Image.fromarray(new_image[:, :, [2, 1, 0]])
    if save_path:
        pil_image.save(save_path)
    if show:
        pil_image.show()

    return new_image


# (The same image from the previous code example)
url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.png"
image = Image.open(requests.get(url, stream=True).raw)

# From the previous code example
entities = [('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a fire', (41, 47), [(0.171875, 0.015625, 0.484375, 0.890625)])]

# Draw the bounding bboxes
draw_entity_boxes_on_image(image, entities, show=True)

Here is the annotated image:

BibTex and citation info

@article{kosmos-2,
  title={Kosmos-2: Grounding Multimodal Large Language Models to the World},
  author={Zhiliang Peng and Wenhui Wang and Li Dong and Yaru Hao and Shaohan Huang and Shuming Ma and Furu Wei},
  journal={ArXiv},
  year={2023},
  volume={abs/2306}
}

@article{kosmos-1,
  title={Language Is Not All You Need: Aligning Perception with Language Models},
  author={Shaohan Huang and Li Dong and Wenhui Wang and Yaru Hao and Saksham Singhal and Shuming Ma and Tengchao Lv and Lei Cui and Owais Khan Mohammed and Qiang Liu and Kriti Aggarwal and Zewen Chi and Johan Bjorck and Vishrav Chaudhary and Subhojit Som and Xia Song and Furu Wei},
  journal={ArXiv},
  year={2023},
  volume={abs/2302.14045}
}

@article{metalm,
  title={Language Models are General-Purpose Interfaces},
  author={Yaru Hao and Haoyu Song and Li Dong and Shaohan Huang and Zewen Chi and Wenhui Wang and Shuming Ma and Furu Wei},
  journal={ArXiv},
  year={2022},
  volume={abs/2206.06336}
}

Configuration

Architecture: Kosmos2ForConditionalGeneration
Context length (tokens): 2,048
Vocabulary size: 65,037
Stored precision: float32
Model type: kosmos-2

Identity and Version

Repository: microsoft/kosmos-2-patch14-224
Publisher: Microsoft
Task: Image to text
Modality: Image and text
Library: transformers
Parameters: 1.7B parameters
Languages: Not stated by the source
Revision: e91cfbcb4ce051b6a55bfb5f96165a3bbf5eb82c
First published: 2023-10-02
Last updated: 2023-11-28

Files and Weights

16 files, 13.3 GB in total. The weights are 2 files totalling 13.3 GB in bin, safetensors.

Weights2 files · 13.3 GB

Configuration5 files · 38.2 KB

Tokenizer2 files · 4.9 MB

Documentation1 file · 15.8 KB

Other5 files · 3.0 MB

Repository1 file · 1.6 KB

Every file

File	Type	Size	SHA-256
model.safetensors	Weights	6.7 GB	051bf4b62a25
pytorch_model.bin	Weights	6.7 GB	47e771fe58dd
added_tokens.json	Configuration	32.0 KB	—
config.json	Configuration	4.5 KB	—
generation_config.json	Configuration	137 B	—
preprocessor_config.json	Configuration	534 B	—
special_tokens_map.json	Configuration	1.1 KB	—
README.md	Documentation	15.8 KB	—
annotated_snowman.jpg	Other	124.0 KB	—
sentencepiece.bpe.model	Other	1.4 MB	3a60b4d1d1d8
snowman.jpg	Other	103.6 KB	—
snowman.png	Other	1.4 MB	b97825997df0
two_dogs.jpg	Other	80.0 KB	—
.gitattributes	Repository	1.6 KB	—
tokenizer.json	Tokenizer	4.7 MB	—
tokenizer_config.json	Tokenizer	190.7 KB	—

License and Download

License: mit
Access: Open weights, no gate
Download size: 13.3 GB

Download from Microsoft

Released by Microsoft through its official repository on Hugging Face. Read the license.

Memory Requirements

Precision	Weights in memory
As published	13.3 GB
16-bit	3.3 GB
8-bit	1.7 GB
4-bit	0.8 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About kosmos-2-patch14-224

How much GPU memory does kosmos-2-patch14-224 need?

About 4 GB at 16-bit and 1 GB at 4-bit: the weights (1.7B parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run kosmos-2-patch14-224 on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use kosmos-2-patch14-224 commercially?

Yes. kosmos-2-patch14-224 is released under MIT License. The MIT License is a short permissive license. It permits commercial use, modification and redistribution, provided the copyright notice and permission notice are included.

What is kosmos-2-patch14-224's context length?

2,048 tokens, from the maximum position embeddings in its published configuration.

Similar Models

Model · Image to text

GLM-OCR

Z.ai

Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance…

Open weights mit 1.3B parameters 131,072 tokens transformers

View model

Model · Image to text

GLM-OCR

Unsloth AI

Join our WeChat and Discord community Use GLM-OCR's API GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder–decoder architecture. It introduces Multi-Token Prediction (MTP) loss and stable full-task reinforcement learning to improve training efficiency, recognition accuracy, and generalization. The model integrates the CogViT visual encoder pre-trained on large-scale image–text data, a lightweight cross-modal connector with efficient token downsampling, and a GLM-0.5B language decoder. Combined with a two-stage pipeline of layout analysis and parallel recognition based on PP-DocLayout-V3, GLM-OCR delivers robust and high-quality OCR performance…

Open weights mit 1.3B parameters 131,072 tokens transformers

View model

Model · Image to text

LightOnOCR-1B-1025

LightOn AI

Full BF16 version of the model. We recommend this variant for inference and further fine-tuning. LightOnOCR-1B is a compact, end-to-end vision–language model for Optical Character Recognition (OCR) and document understanding. It achieves state-of-the-art accuracy in its weight class while being several times faster and cheaper than larger general-purpose VLMs. Highlights LightOnOCR combines a Vision Transformer encoder(Pixtral-based) with a lightweight text decoder(Qwen3-based) distilled from high-quality open VLMs. It is optimized for document parsing tasks, producing accurate, layout-aware text extraction from high-resolution pages. All benchmarks evaluated using vLLM on the Olmo-Bench.…

Open weights apache-2.0 1.2B parameters 8,192 tokens transformers

View model

Model · Image to text

GLM-OCR-4bit

MLX Community

This model was converted to MLX format from zai-org/GLM-OCR using mlx-vlm version 0.3.10. Refer to the original model card for more details on the model.

Open weights mit 1.1B parameters 131,072 tokens transformers

View model

Model · Image to text

trocr-large-printed

Microsoft

TrOCR model fine-tuned on the SROIE dataset. It was introduced in the paper TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models by Li et al. and first released in this repository. Disclaimer: The team releasing TrOCR did not write a model card for this model so this model card has been written by the Hugging Face team. The TrOCR model is an encoder-decoder model, consisting of an image Transformer as encoder, and a text Transformer as decoder. The image encoder was initialized from the weights of BEiT, while the text decoder was initialized from the weights of RoBERTa. Images are presented to the model as a sequence of fixed-size patches (resolution 16x16), which…

Open weights 608M parameters transformers

View model

Model · Image to text

blip-image-captioning-large

Salesforce AI Research

captioning pretrained on COCO dataset - base architecture (with ViT large backbone). Authors from the paper write in the abstract: Vision-Language Pre-training (VLP) has advanced the performance for many vision-language tasks. However, most existing pre-trained models only excel in either understanding-based tasks or generation-based tasks. Furthermore, performance improvement has been largely achieved by scaling up the dataset with noisy image-text pairs collected from the web, which is a suboptimal source of supervision. In this paper, we propose BLIP, a new VLP framework which transfers flexibly to both vision-language understanding and generation tasks. BLIP effectively utilizes the…

Open weights bsd-3-clause 470M parameters 512 tokens transformers

View model