openbmb/MiniCPM-V-4_5

Key Techniques

• Architechture: Unified 3D-Resampler for High-density Video Compression. MiniCPM-V 4.5 introduces a 3D-Resampler that overcomes the performance-efficiency trade-off in video understanding. By grouping and jointly compressing up to 6 consecutive video frames into just 64 tokens (the same token count used for a single image in MiniCPM-V series), MiniCPM-V 4.5 achieves a 96× compression rate for video tokens. This allows the model to process more video frames without additional LLM computational cost, enabling high-FPS video and long video understanding. The architecture supports unified encoding for images, multi-image inputs, and videos, ensuring seamless capability and knowledge transfer.

• Pre-training: Unified Learning for OCR and Knowledge from Documents. Existing MLLMs learn OCR capability and knowledge from documents in isolated training approaches. We observe that the essential difference between these two training approaches is the visibility of the text in images. By dynamically corrupting text regions in documents with varying noise levels and asking the model to reconstruct the text, the model learns to adaptively and properly switch between accurate text recognition (when text is visible) and multimodal context-based knowledge reasoning (when text is heavily obscured). This eliminates reliance on error-prone document parsers in knowledge learning from documents, and prevents hallucinations from over-augmented OCR data, resulting in top-tier OCR and multimodal knowledge performance with minimal engineering overhead.

• Post-training: Hybrid Fast/Deep Thinking with Multimodal RL. MiniCPM-V 4.5 offers a balanced reasoning experience through two switchable modes: fast thinking for efficient daily use and deep thinking for complex tasks. Using a new hybrid reinforcement learning method, the model jointly optimizes both modes, significantly enhancing fast-mode performance without compromising deep-mode capability. Incorporated with RLPR and RLAIF-V, it generalizes robust reasoning skills from broad multimodal data while effectively reducing hallucinations.

Evaluation

Inference Efficiency

OpenCompass

Video-MME

Both Video-MME and OpenCompass were evaluated using 8×A100 GPUs for inference. The reported inference time of Video-MME includes full model-side computation, and excludes the external cost of video frame extraction (dependent on specific frame extraction tools) for fair comparison.

Examples

We deploy MiniCPM-V 4.5 on iPad M4 with iOS demo. The demo video is the raw screen recording without editing.

Framework Support Matrix

Note: If you'd like us to prioritize support for another open-source framework, please let us know via this short form.

Usage

If you wish to enable thinking mode, provide the argument enable_thinking=True to the chat function.

Chat with Image

import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

torch.manual_seed(100)

model = AutoModel.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True, # or openbmb/MiniCPM-o-2_6
    attn_implementation='sdpa', torch_dtype=torch.bfloat16) # sdpa or flash_attention_2, no eager
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True) # or openbmb/MiniCPM-o-2_6

image = Image.open('./assets/minicpmo2_6/show_demo.jpg').convert('RGB')

enable_thinking=False # If `enable_thinking=True`, the thinking mode is enabled.
stream=True # If `stream=True`, the answer is string

# First round chat 
question = "What is the landform in the picture?"
msgs = [{'role': 'user', 'content': [image, question]}]

answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    enable_thinking=enable_thinking,
    stream=True
)

generated_text = ""
for new_text in answer:
    generated_text += new_text
    print(new_text, flush=True, end='')

# Second round chat, pass history context of multi-turn conversation
msgs.append({"role": "assistant", "content": [generated_text]})
msgs.append({"role": "user", "content": ["What should I pay attention to when traveling here?"]})

answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    stream=True
)

generated_text = ""
for new_text in answer:
    generated_text += new_text
    print(new_text, flush=True, end='')

You will get the following output:

# round1
The landform in the picture is karst topography. Karst landscapes are characterized by distinctive, jagged limestone hills or mountains with steep, irregular peaks and deep valleys—exactly what you see here These unique formations result from the dissolution of soluble rocks like limestone over millions of years through water erosion.

This scene closely resembles the famous karst landscape of Guilin and Yangshuo in China’s Guangxi Province. The area features dramatic, pointed limestone peaks rising dramatically above serene rivers and lush green forests, creating a breathtaking and iconic natural beauty that attracts millions of visitors each year for its picturesque views.

# round2
When traveling to a karst landscape like this, here are some important tips:

1. Wear comfortable shoes: The terrain can be uneven and hilly.
2. Bring water and snacks for energy during hikes or boat rides.
3. Protect yourself from the sun with sunscreen, hats, and sunglasses—especially since you’ll likely spend time outdoors exploring scenic spots.
4. Respect local customs and nature regulations by not littering or disturbing wildlife.

By following these guidelines, you'll have a safe and enjoyable trip while appreciating the stunning natural beauty of places such as Guilin’s karst mountains.

Chat with Video

## The 3d-resampler compresses multiple frames into 64 tokens by introducing temporal_ids. 
# To achieve this, you need to organize your video data into two corresponding sequences: 
#   frames: List[Image]
#   temporal_ids: List[List[Int]].

import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer
from decord import VideoReader, cpu    # pip install decord
from scipy.spatial import cKDTree
import numpy as np
import math

model = AutoModel.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True,  # or openbmb/MiniCPM-o-2_6
    attn_implementation='sdpa', torch_dtype=torch.bfloat16) # sdpa or flash_attention_2, no eager
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True)  # or openbmb/MiniCPM-o-2_6

MAX_NUM_FRAMES=180 # Indicates the maximum number of frames received after the videos are packed. The actual maximum number of valid frames is MAX_NUM_FRAMES * MAX_NUM_PACKING.
MAX_NUM_PACKING=3  # indicates the maximum packing number of video frames. valid range: 1-6
TIME_SCALE = 0.1 

def map_to_nearest_scale(values, scale):
    tree = cKDTree(np.asarray(scale)[:, None])
    _, indices = tree.query(np.asarray(values)[:, None])
    return np.asarray(scale)[indices]


def group_array(arr, size):
    return [arr[i:i+size] for i in range(0, len(arr), size)]

def encode_video(video_path, choose_fps=3, force_packing=None):
    def uniform_sample(l, n):
        gap = len(l) / n
        idxs = [int(i * gap + gap / 2) for i in range(n)]
        return [l[i] for i in idxs]
    vr = VideoReader(video_path, ctx=cpu(0))
    fps = vr.get_avg_fps()
    video_duration = len(vr) / fps
        
    if choose_fps * int(video_duration) <= MAX_NUM_FRAMES:
        packing_nums = 1
        choose_frames = round(min(choose_fps, round(fps)) * min(MAX_NUM_FRAMES, video_duration))
        
    else:
        packing_nums = math.ceil(video_duration * choose_fps / MAX_NUM_FRAMES)
        if packing_nums <= MAX_NUM_PACKING:
            choose_frames = round(video_duration * choose_fps)
        else:
            choose_frames = round(MAX_NUM_FRAMES * MAX_NUM_PACKING)
            packing_nums = MAX_NUM_PACKING

    frame_idx = [i for i in range(0, len(vr))]      
    frame_idx =  np.array(uniform_sample(frame_idx, choose_frames))

    if force_packing:
        packing_nums = min(force_packing, MAX_NUM_PACKING)
    
    print(video_path, ' duration:', video_duration)
    print(f'get video frames={len(frame_idx)}, packing_nums={packing_nums}')
    
    frames = vr.get_batch(frame_idx).asnumpy()

    frame_idx_ts = frame_idx / fps
    scale = np.arange(0, video_duration, TIME_SCALE)

    frame_ts_id = map_to_nearest_scale(frame_idx_ts, scale) / TIME_SCALE
    frame_ts_id = frame_ts_id.astype(np.int32)

    assert len(frames) == len(frame_ts_id)

    frames = [Image.fromarray(v.astype('uint8')).convert('RGB') for v in frames]
    frame_ts_id_group = group_array(frame_ts_id, packing_nums)
    
    return frames, frame_ts_id_group


video_path="video_test.mp4"
fps = 5 # fps for video
force_packing = None # You can set force_packing to ensure that 3D packing is forcibly enabled; otherwise, encode_video will dynamically set the packing quantity based on the duration.
frames, frame_ts_id_group = encode_video(video_path, fps, force_packing=force_packing)

question = "Describe the video"
msgs = [
    {'role': 'user', 'content': frames + [question]}, 
]


answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    use_image_id=False,
    max_slice_nums=1,
    temporal_ids=frame_ts_id_group
)
print(answer)

Chat with multiple images

import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True,
    attn_implementation='sdpa', torch_dtype=torch.bfloat16) # sdpa or flash_attention_2
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True)

image1 = Image.open('image1.jpg').convert('RGB')
image2 = Image.open('image2.jpg').convert('RGB')
question = 'Compare image 1 and image 2, tell me about the differences between image 1 and image 2.'

msgs = [{'role': 'user', 'content': [image1, image2, question]}]

answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer
)
print(answer)

In-context few-shot learning

import torch
from PIL import Image
from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True,
    attn_implementation='sdpa', torch_dtype=torch.bfloat16)
model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-V-4_5', trust_remote_code=True)

question = "production date" 
image1 = Image.open('example1.jpg').convert('RGB')
answer1 = "2023.08.04"
image2 = Image.open('example2.jpg').convert('RGB')
answer2 = "2007.04.24"
image_test = Image.open('test.jpg').convert('RGB')

msgs = [
    {'role': 'user', 'content': [image1, question]}, {'role': 'assistant', 'content': [answer1]},
    {'role': 'user', 'content': [image2, question]}, {'role': 'assistant', 'content': [answer2]},
    {'role': 'user', 'content': [image_test, question]}
]

answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer
)
print(answer)

License

Model License

• The MiniCPM-o/V model weights and code are open-sourced under the Apache-2.0 license.
• To help us better understand and support our users, we would deeply appreciate it if you could consider optionally filling out a brief registration "questionnaire".

Statement

• As an LMM, MiniCPM-V 4.5 generates contents by learning a large amount of multimodal corpora, but it cannot comprehend, express personal opinions or make value judgement. Anything generated by MiniCPM-V 4.5 does not represent the views and positions of the model developers
• We will not be liable for any problems arising from the use of the MinCPM-V models, including but not limited to data security issues, risk of public opinion, or any risks and problems arising from the misdirection, misuse, dissemination or misuse of the model.

Key Techniques and Other Multimodal Projects

👏 Welcome to explore key techniques of MiniCPM-V 4.5 and other multimodal projects of our team:

VisCPM | RLPR | RLHF-V | LLaVA-UHD | RLAIF-V

Citation

If you find our work helpful, please consider citing our papers 📝 and liking this project ❤️！

@misc{yu2025minicpmv45cookingefficient,
      title={MiniCPM-V 4.5: Cooking Efficient MLLMs via Architecture, Data, and Training Recipe}, 
      author={Tianyu Yu and Zefan Wang and Chongyi Wang and Fuwei Huang and Wenshuo Ma and Zhihui He and Tianchi Cai and Weize Chen and Yuxiang Huang and Yuanqian Zhao and Bokai Xu and Junbo Cui and Yingjing Xu and Liqing Ruan and Luoyuan Zhang and Hanyu Liu and Jingkun Tang and Hongyuan Liu and Qining Guo and Wenhao Hu and Bingxiang He and Jie Zhou and Jie Cai and Ji Qi and Zonghao Guo and Chi Chen and Guoyang Zeng and Yuxuan Li and Ganqu Cui and Ning Ding and Xu Han and Yuan Yao and Zhiyuan Liu and Maosong Sun},
      year={2025},
      eprint={2509.18154},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.18154}, 
}

@article{yao2024minicpm,
  title={MiniCPM-V: A GPT-4V Level MLLM on Your Phone},
  author={Yao, Yuan and Yu, Tianyu and Zhang, Ao and Wang, Chongyi and Cui, Junbo and Zhu, Hongji and Cai, Tianchi and Li, Haoyu and Zhao, Weilin and He, Zhihui and others},
  journal={Nat Commun 16, 5509 (2025)},
  year={2025}
}