端侧可用的 GPT-4o 级视觉、语音、多模态流式大模型

GitHub | Online Demo

MiniCPM-o 2.6 是 MiniCPM-o 系列的最新、性能最佳模型。该模型基于 SigLip-400M、Whisper-medium-300M、ChatTTS-200M 和 Qwen2.5-7B 构建，共 8B 参数，通过端到端方式训练和推理。相比 MiniCPM-V 2.6，该模型在性能上有了显著提升，并支持了实时语音对话和多模态流式交互的新功能。MiniCPM-o 2.6 的主要特性包括：

• 🔥 领先的视觉能力。 MiniCPM-o 2.6 在 OpenCompass 榜单上（综合 8 个主流多模态评测基准）平均得分 70.2，以 8B 量级的大小在单图理解方面超越了 GPT-4o-202405、Gemini 1.5 Pro 和 Claude 3.5 Sonnet 等主流商用闭源多模态大模型。此外，它的多图和视频理解表现也优于 GPT-4V 和 Claude 3.5 Sonnet，并展现出了优秀的上下文学习能力。

• 🎙 出色的语音能力。 MiniCPM-o 2.6 支持可配置声音的中英双语实时对话。MiniCPM-o 2.6 在语音理解任务（如 ASR 和 STT translation）上的表现优于 GPT-4o-realtime，并在语音对话的语义和声学评估中展现了开源模型中最高的语音生成性能。它还支持情绪/语速/风格控制、语音克隆、角色扮演等进阶能力。

• 🎬 强大的多模态流式交互能力。 作为一项新功能，MiniCPM-o 2.6 能够接受连续的视频和音频流，并和用户进行实时语音交互。在 StreamingBench（针对实时视频理解、全模态（视/音频）理解、多模态上下文理解的综合评测基准）中，MiniCPM-o 2.6 获得开源模型最高分并超过了 GPT-4o-realtime 和 Claude 3.5 Sonnet。

• 💪 强大的 OCR 能力及其他功能。 MiniCPM-o 2.6 进一步优化了 MiniCPM-V 2.6 的众多视觉理解能力，其可以处理任意长宽比的图像，像素数可达 180 万（如 1344x1344）。在 OCRBench 上取得25B 以下最佳水平，超过 GPT-4o-202405 等商用闭源模型。基于最新的 RLHF-V、RLAIF-V 和 VisCPM 技术，其具备了可信的多模态行为，在 MMHal-Bench 上超过了 GPT-4o 和 Claude 3.5，并支持英语、中文、德语、法语、意大利语、韩语等多种语言。

• 🚀 卓越的效率。 除了对个人用户友好的模型大小，MiniCPM-o 2.6 还表现出最先进的视觉 token 密度（即每个视觉 token 编码的像素数量）。它仅需 640 个 token 即可处理 180 万像素图像，比大多数模型少 75%。这一特性优化了模型的推理速度、首 token 延迟、内存占用和功耗。因此，MiniCPM-o 2.6 可以支持 iPad 等终端设备上的高效多模态流式交互。

• 💫 易于使用。 MiniCPM-o 2.6 可以通过多种方式轻松使用：(1) llama.cpp 支持在本地设备上进行高效的 CPU 推理，(2) int4 和 GGUF 格式的量化模型，有 16 种尺寸，(3) vLLM 支持高吞吐量和内存高效的推理，(4) 通过LLaMA-Factory框架针对新领域和任务进行微调，(5) 使用 Gradio 快速设置本地 WebUI 演示，(6) 在线demo。

模型架构。

• 端到端全模态架构。 通过端到端的方式连接和训练不同模态的编/解码模块以充分利用丰富的多模态知识。
• 全模态流式机制。 (1) 我们将不同模态的离线编/解码器改造为适用于流式输入/输出的在线模块。 (2) 我们针对大语言模型基座设计了一种时分复用的全模态流式信息处理机制，将平行的不同模态的信息流拆分重组为周期性时间片序列。
• 可配置的声音方案。 我们设计了包含传统文本系统提示词和用于指定模型声音的语音系统提示词结构。从而，模型可在推理时灵活地通过文字或语音样例控制声音风格，支持声音克隆和声音生成等高级能力。

以下示例为 MiniCPM-o 2.6 部署在 iPad Pro 和web demo录制得到。

Inference using Huggingface transformers on NVIDIA GPUs. Please ensure that transformers==4.44.2 is installed, as other versions may have compatibility issues. We are investigating this issue. Requirements tested on python 3.10：

Pillow==10.1.0
torch==2.3.1
torchaudio==2.3.1
torchvision==0.18.1
transformers==4.44.2
librosa==0.9.0
soundfile==0.12.1
vector-quantize-pytorch==1.18.5
vocos==0.1.0
decord
moviepy

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

# load omni model default, the default init_vision/init_audio/init_tts is True
# if load vision-only model, please set init_audio=False and init_tts=False
# if load audio-only model, please set init_vision=False
model = AutoModel.from_pretrained(
    'openbmb/MiniCPM-o-2_6',
    trust_remote_code=True,
    attn_implementation='sdpa', # sdpa or flash_attention_2
    torch_dtype=torch.bfloat16,
    init_vision=True,
    init_audio=True,
    init_tts=True
)


model = model.eval().cuda()
tokenizer = AutoTokenizer.from_pretrained('openbmb/MiniCPM-o-2_6', trust_remote_code=True)

# In addition to vision-only mode, tts processor and vocos also needs to be initialized
model.init_tts()

If you are using an older version of PyTorch, you might encounter this issue "weight_norm_fwd_first_dim_kernel" not implemented for 'BFloat16', Please convert the TTS to float32 type.

model.tts.float()

We provide two inference modes: chat and streaming

import math
import numpy as np
from PIL import Image
from moviepy.editor import VideoFileClip
import tempfile
import librosa
import soundfile as sf

def get_video_chunk_content(video_path, flatten=True):
    video = VideoFileClip(video_path)
    print('video_duration:', video.duration)
    
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=True) as temp_audio_file:
        temp_audio_file_path = temp_audio_file.name
        video.audio.write_audiofile(temp_audio_file_path, codec="pcm_s16le", fps=16000)
        audio_np, sr = librosa.load(temp_audio_file_path, sr=16000, mono=True)
    num_units = math.ceil(video.duration)
    
    # 1 frame + 1s audio chunk
    contents= []
    for i in range(num_units):
        frame = video.get_frame(i+1)
        image = Image.fromarray((frame).astype(np.uint8))
        audio = audio_np[sr*i:sr*(i+1)]
        if flatten:
            contents.extend(["<unit>", image, audio])
        else:
            contents.append(["<unit>", image, audio])
    
    return contents

video_path="assets/Skiing.mp4"
# if use voice clone prompt, please set ref_audio
ref_audio_path = 'assets/demo.wav'
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
sys_msg = model.get_sys_prompt(ref_audio=ref_audio, mode='omni', language='en')
# or use default prompt
# sys_msg = model.get_sys_prompt(mode='omni', language='en')

contents = get_video_chunk_content(video_path)
msg = {"role":"user", "content": contents}
msgs = [sys_msg, msg]

# please set generate_audio=True and output_audio_path to save the tts result
generate_audio = True
output_audio_path = 'output.wav'

res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    temperature=0.5,
    max_new_tokens=4096,
    omni_input=True, # please set omni_input=True when omni inference
    use_tts_template=True,
    generate_audio=generate_audio,
    output_audio_path=output_audio_path,
    max_slice_nums=1,
    use_image_id=False,
    return_dict=True
)
print(res)

## You will get the answer: The person in the picture is skiing down a snowy slope.
# import IPython
# IPython.display.Audio('output.wav')

# a new conversation need reset session first, it will reset the kv-cache
model.reset_session()

contents = get_video_chunk_content(video_path, flatten=False)
session_id = '123'
generate_audio = True

# 1. prefill system prompt
res = model.streaming_prefill(
    session_id=session_id,
    msgs=[sys_msg], 
    tokenizer=tokenizer
)

# 2. prefill video/audio chunks
for content in contents:
    msgs = [{"role":"user", "content": content}]
    res = model.streaming_prefill(
        session_id=session_id,
        msgs=msgs, 
        tokenizer=tokenizer
    )

# 3. generate
res = model.streaming_generate(
    session_id=session_id,
    tokenizer=tokenizer,
    temperature=0.5,
    generate_audio=generate_audio
)

audios = []
text = ""

if generate_audio:
    for r in res:
        audio_wav = r.audio_wav
        sampling_rate = r.sampling_rate
        txt = r.text

        audios.append(audio_wav)
        text += txt
        
    res = np.concatenate(audios)
    sf.write("output.wav", res, samplerate=sampling_rate)
    print("text:", text)
    print("audio saved to output.wav")
else:
    for r in res:
        text += r['text']
    print("text:", text)

Model initialization

import torch
import librosa
from transformers import AutoModel, AutoTokenizer

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

model.init_tts()
model.tts.float()

Mimick task reflects a model's end-to-end speech modeling capability. The model takes audio input, and outputs an ASR transcription and subsequently reconstructs the original audio with high similarity. The higher the similarity between the reconstructed audio and the original audio, the stronger the model's foundational capability in end-to-end speech modeling.

mimick_prompt = "Please repeat each user's speech, including voice style and speech content."
audio_input, _ = librosa.load('./assets/input_examples/Trump_WEF_2018_10s.mp3', sr=16000, mono=True) # load the audio to be mimicked

# can also try `./assets/input_examples/cxk_original.wav`, 
# `./assets/input_examples/fast-pace.wav`, 
# `./assets/input_examples/chi-english-1.wav` 
# `./assets/input_examples/exciting-emotion.wav` 
# for different aspects of speech-centric features.

msgs = [{'role': 'user', 'content': [mimick_prompt, audio_input]}]
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    temperature=0.3,
    generate_audio=True,
    output_audio_path='output_mimick.wav', # save the tts result to output_audio_path
)

A general usage scenario of MiniCPM-o-2.6 is role-playing a specific character based on the audio prompt. It will mimic the voice of the character to some extent and act like the character in text, including language style. In this mode, MiniCPM-o-2.6 sounds more natural and human-like. Self-defined audio prompts can be used to customize the voice of the character in an end-to-end manner.

ref_audio, _ = librosa.load('./assets/input_examples/icl_20.wav', sr=16000, mono=True) # load the reference audio
sys_prompt = model.get_sys_prompt(ref_audio=ref_audio, mode='audio_roleplay', language='en')

# round one
user_question = {'role': 'user', 'content': [librosa.load('xxx.wav', sr=16000, mono=True)[0]]}
msgs = [sys_prompt, user_question]
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_roleplay_round_1.wav',
)

# round two
history = msgs.append({'role': 'assistant', 'content': res})
user_question = {'role': 'user', 'content': [librosa.load('xxx.wav', sr=16000, mono=True)[0]]}
msgs = history.append(user_question)
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_roleplay_round_2.wav',
)
print(res)

An enhanced feature of MiniCPM-o-2.6 is to act as an AI assistant, but only with limited choice of voices. In this mode, MiniCPM-o-2.6 is less human-like and more like a voice assistant. In this mode, the model is more instruction-following. For demo, you are suggested to use assistant_default_female_voice, assistant_male_voice. Other voices may work but not as stable as the default voices.

ref_audio, _ = librosa.load('./assets/input_examples/assistant_default_female_voice.wav', sr=16000, mono=True) # or use `./assets/input_examples/assistant_male_voice.wav`
sys_prompt = model.get_sys_prompt(ref_audio=ref_audio, mode='audio_assistant', language='en') 
user_question = {'role': 'user', 'content': [librosa.load('xxx.wav', sr=16000, mono=True)[0]]} # load the user's audio question

# round one
msgs = [sys_prompt, user_question]
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_assistant_round_1.wav',
)

# round two
history = msgs.append({'role': 'assistant', 'content': res})
user_question = {'role': 'user', 'content': [librosa.load('xxx.wav', sr=16000, mono=True)[0]]}
msgs = history.append(user_question)
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_assistant_round_2.wav',
)
print(res)

MiniCPM-o-2.6 can also do Instruction-to-Speech, aka Voice Creation. You can describe a voice in detail, and the model will generate a voice that matches the description. For more Instruction-to-Speech sample instructions, you can refer to https://voxinstruct.github.io/VoxInstruct/.

instruction = 'Speak like a male charming superstar, radiating confidence and style in every word.'

msgs = [{'role': 'user', 'content': [instruction]}]

res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_voice_creation.wav',
)

MiniCPM-o-2.6 can also do zero-shot text-to-speech, aka Voice Cloning. With this mode, model will act like a TTS model.

ref_audio, _ = librosa.load('./assets/input_examples/icl_20.wav', sr=16000, mono=True) # load the reference audio
sys_prompt = model.get_sys_prompt(ref_audio=ref_audio, mode='voice_cloning', language='en')
text_prompt = f"Please read the text below."
user_question = {'role': 'user', 'content': [text_prompt, "content that you want to read"]}

msgs = [sys_prompt, user_question]
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_voice_cloning.wav',
)

MiniCPM-o-2.6 can also be used to address various audio understanding tasks, such as ASR, speaker analysis, general audio captioning, and sound scene tagging.

For audio-to-text tasks, you can use the following prompts:

• ASR with ZH(same as AST en2zh): 请仔细听这段音频片段，并将其内容逐字记录。
• ASR with EN(same as AST zh2en): Please listen to the audio snippet carefully and transcribe the content.
• Speaker Analysis: Based on the speaker's content, speculate on their gender, condition, age range, and health status.
• General Audio Caption: Summarize the main content of the audio.
• General Sound Scene Tagging: Utilize one keyword to convey the audio's content or the associated scene.

task_prompt = "Please listen to the audio snippet carefully and transcribe the content." + "\n" # can change to other prompts.
audio_input, _ = librosa.load('./assets/input_examples/audio_understanding.mp3', sr=16000, mono=True) # load the audio to be captioned

msgs = [{'role': 'user', 'content': [task_prompt, audio_input]}]

res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    max_new_tokens=128,
    use_tts_template=True,
    generate_audio=True,
    temperature=0.3,
    output_audio_path='result_audio_understanding.wav',
)
print(res)

MiniCPM-o-2_6 has the same inference methods as MiniCPM-V-2_6

# test.py
image = Image.open('xx.jpg').convert('RGB')
question = 'What is in the image?'
msgs = [{'role': 'user', 'content': [image, question]}]
res = model.chat(
    image=None,
    msgs=msgs,
    tokenizer=tokenizer
)
print(res)

## if you want to use streaming, please make sure sampling=True and stream=True
## the model.chat will return a generator
res = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    sampling=True,
    stream=True
)
generated_text = ""
for new_text in res:
    generated_text += new_text
    print(new_text, flush=True, end='')

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)

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)

MAX_NUM_FRAMES=64 # if cuda OOM set a smaller number
def encode_video(video_path):
    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))
    sample_fps = round(vr.get_avg_fps() / 1)  # FPS
    frame_idx = [i for i in range(0, len(vr), sample_fps)]
    if len(frame_idx) > MAX_NUM_FRAMES:
        frame_idx = uniform_sample(frame_idx, MAX_NUM_FRAMES)
    frames = vr.get_batch(frame_idx).asnumpy()
    frames = [Image.fromarray(v.astype('uint8')) for v in frames]
    print('num frames:', len(frames))
    return frames
video_path ="video_test.mp4"
frames = encode_video(video_path)
question = "Describe the video"
msgs = [
    {'role': 'user', 'content': frames + [question]}, 
]
# Set decode params for video
params={}
params["use_image_id"] = False
params["max_slice_nums"] = 2 # use 1 if cuda OOM and video resolution >  448*448
answer = model.chat(
    msgs=msgs,
    tokenizer=tokenizer,
    **params
)
print(answer)

Please look at GitHub for more detail about usage.

MiniCPM-o 2.6 (vision-only mode) can run with llama.cpp. See our fork of llama.cpp and readme for more detail.

int4 量化版，更低的显存占用(9GB): MiniCPM-o-2_6-int4.

• The code in this repo is released under the Apache-2.0 License.
• The usage of MiniCPM-o and MiniCPM-V series model weights must strictly follow MiniCPM Model License.md.
• The models and weights of MiniCPM are completely free for academic research. After filling out a "questionnaire" for registration, MiniCPM-o 2.6 weights are also available for free commercial use.

• As an LMM, MiniCPM-o 2.6 generates contents by learning a large mount of multimodal corpora, but it cannot comprehend, express personal opinions or make value judgement. Anything generated by MiniCPM-o 2.6 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.

👏 Welcome to explore key techniques of MiniCPM-o 2.6 and other multimodal projects of our team:

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

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

@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={arXiv preprint arXiv:2408.01800},
  year={2024}
}