MiniCPM-o 4.5 is the latest and most capable model in the MiniCPM-o series. The model is built in an end-to-end fashion based on SigLip2, Whisper-medium, CosyVoice2, and Qwen3-8B with a total of 9B parameters. It exhibits a significant performance improvement, and introduces new features for full-duplex multimodal live streaming. Notable features of MiniCPM-o 4.5 include:
🔥 Leading Visual Capability. MiniCPM-o 4.5 achieves an average score of 77.6 on OpenCompass, a comprehensive evaluation of 8 popular benchmarks. With only 9B parameters, it surpasses widely used proprietary models like GPT-4o, Gemini 2.0 Pro, and approaches Gemini 2.5 Flash for vision-language capabilities. It supports instruct and thinking modes in a single model, better covering efficiency and performance trade-offs in different user scenarios.
🎙 Strong Speech Capability. MiniCPM-o 4.5 supports bilingual real-time speech conversation with configurable voices in English and Chinese. It features more natural, expressive and stable speech conversation. The model also allows for fun features such as voice cloning and role play via a simple reference audio clip, where the cloning performance surpasses strong TTS tools such as CosyVoice2.
🎬 New Full-Duplex and Proactive Multimodal Live Streaming Capability. As a new feature, MiniCPM-o 4.5 can process real-time, continuous video and audio input streams simultaneously while generating concurrent text and speech output streams in an end-to-end fashion, without mutual blocking. This allows MiniCPM-o 4.5 to see, listen, and speak simultaneously, creating a fluid, real-time omnimodal conversation experience. Beyond reactive responses, the model can also perform proactive interaction, such as initiating reminders or comments based on its continuous understanding of the live scene.
💪 Strong OCR Capability, Efficiency and Others.
Advancing popular visual capabilities from MiniCPM-V series, MiniCPM-o 4.5 can process high-resolution images (up to 1.8 million pixels) and high-FPS videos (up to 10fps) in any aspect ratio efficiently. It achieves state-of-the-art performance for end-to-end English document parsing on OmniDocBench, outperforming proprietary models such as Gemini-3 Flash and GPT-5, and specialized tools such as DeepSeek-OCR 2. It also features trustworthy behaviors, matching Gemini 2.5 Flash on MMHal-Bench, and supports multilingual capabilities on more than 30 languages.
💫 Easy Usage. MiniCPM-o 4.5 can be easily used in various ways: Basic usage, recommended for 100% precision: PyTorch inference with Nvidia GPU. Other end-side adaptation includes (1) llama.cpp and Ollama support for efficient CPU inference on local devices, (2) int4 and GGUF format quantized models in 16 sizes, (3) vLLM and SGLang support for high-throughput and memory-efficient inference, (4) FlagOS support for the unified multi-chip backend plugin. We also open-sourced web demos on which enables the full-duplex multimodal live streaming experience on local devices such as GPUs, PCs (e.g., on a MacBook).
Model Architecture.
End-to-end Omni-modal Architecture. The modality encoders/decoders and LLM are densely connected via hidden states in an end-to-end fashion. This enables better information flow and control, and also facilitates full exploitation of rich multimodal knowledge during training.
Full-Duplex Omni-modal Live Streaming Mechanism. (1) We turn the offline modality encoder/decoders into online and full-duplex ones for streaming inputs/outputs. The speech token decoder models text and speech tokens in an interleaved fashion to support full-duplex speech generation (i.e., sync timely with new input). This also facilitates more stable long speech generation (e.g., > 1min).
(2) We sync all the input and output streams on timeline in milliseconds, which are jointly modeled by a time-division multiplexing (TDM) mechanism for omni-modality streaming processing in the LLM backbone. It divides parallel omni-modality streams into sequential info groups within small periodic time slices.
Proactive Interaction Mechanism. The LLM continuously monitors the input video and audio streams, and decides at a frequency of 1Hz to speak or not. This high decision-making frequency together with full-duplex nature are curcial to enable the proactive interaction capability.
Configurable Speech Modeling Design. We inherent the multimodal system prompt design of MiniCPM-o 2.6, which includes a traditional text system prompt, and a new audio system prompt to determine the assistant voice. This enables cloning new voices and role play in inference time for speech conversation.
Note: Scores marked with ∗ are from our evaluation; others are cited from referenced reports. n/a indicates that the model does not support the corresponding modality. All results are reported in instruct mode/variant.
Note. For detailed speech conversation examples, refer to Audio Casebook
🚀 Elon Musk - Voice Roleplay (EN)
| System: Clone the voice in the provided audio prompt.
🎙️ [Reference Voice] ➡️
Please chat with the user in a highly human-like and oral style. You are Elon Musk, CEO of Tesla and SpaceX. You speak directly and casually, often with dry humor. You're passionate about Mars, sustainable energy, and pushing humanity forward. Speak bluntly with occasional dark humor. Use simple logic and don't sugarcoat things. Don't be diplomatic. Say what you actually think, even if it's controversial. Keep responses around 100 words. Don't ramble. |
| User: 💬/🎙️ Elon, what is the real reason that you want to go to Mars?
| AI Response: 💬+🎙️
I think it's important for humanity to become a multi-planetary species. I mean, we're sort of at the point where if there was some kind of global catastrophe or something like that, then we'd be in big trouble. So by becoming a multi-planetary species, I just want us to sort of guarantee our future as humans. And Mars is obviously the obvious place to go because it has similar conditions to Earth and it also has water on its surface. So yeah, basically I'm trying to save humanity from itself. |
Inference using Hugging Face Transformers on NVIDIA GPUs. Please ensure transformers==4.51.0 is installed, as other versions may have compatibility issues (under investigation). Requirements tested on Python 3.10:
Note: FFmpeg is required for video frame extraction (get_video_frame_audio_segments with use_ffmpeg=True) and video generation (generate_duplex_video). For more information, visit the official FFmpeg website.
import torch
from transformers import AutoModel
# Load omni model (default: init_vision=True, init_audio=True, init_tts=True)
# For vision-only model: set init_audio=False and init_tts=False
# For audio-only model: set init_vision=False
model = AutoModel.from_pretrained(
"openbmb/MiniCPM-o-4_5",
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.eval().cuda()
# Initialize TTS for audio output
model.init_tts()
# Convert half-duplex model to duplex mode
duplex_model = model.as_duplex()
# Convert duplex model back to half-duplex mode
model = duplex_model.as_simplex(reset_session=True)
Full-duplex streaming inference for real-time or recorded video conversations.
import librosa
import torch
from minicpmo.utils import generate_duplex_video, get_video_frame_audio_segments
from transformers import AutoModel
# Load model and convert to duplex mode
model = AutoModel.from_pretrained(
"openbmb/MiniCPM-o-4_5",
trust_remote_code=True,
attn_implementation="sdpa", # or "flash_attention_2"
torch_dtype=torch.bfloat16,
)
model.eval().cuda()
model = model.as_duplex()
# Load video and reference audio
video_path = "assets/omni_duplex1.mp4"
ref_audio_path = "assets/HT_ref_audio.wav"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
# Extract video frames and audio segments
video_frames, audio_segments, stacked_frames = get_video_frame_audio_segments(
video_path, stack_frames=1, use_ffmpeg=True, adjust_audio_length=True
)
# Prepare duplex session with system prompt and voice reference
model.prepare(
prefix_system_prompt="Streaming Omni Conversation.",
ref_audio=ref_audio,
prompt_wav_path=ref_audio_path,
)
results_log = []
timed_output_audio = []
# Process each chunk in streaming fashion
for chunk_idx in range(len(audio_segments)):
audio_chunk = audio_segments[chunk_idx] if chunk_idx < len(audio_segments) else None
frame = video_frames[chunk_idx] if chunk_idx < len(video_frames) else None
frame_list = []
if frame is not None:
frame_list.append(frame)
if stacked_frames is not None and chunk_idx < len(stacked_frames) and stacked_frames[chunk_idx] is not None:
frame_list.append(stacked_frames[chunk_idx])
# Step 1: Streaming prefill
model.streaming_prefill(
audio_waveform=audio_chunk,
frame_list=frame_list,
max_slice_nums=1, # Increase for HD mode (e.g., [2, 1] for stacked frames)
batch_vision_feed=False, # Set True for faster processing
)
# Step 2: Streaming generate
result = model.streaming_generate(
prompt_wav_path=ref_audio_path,
max_new_speak_tokens_per_chunk=20,
decode_mode="sampling",
)
if result["audio_waveform"] is not None:
timed_output_audio.append((chunk_idx, result["audio_waveform"]))
chunk_result = {
"chunk_idx": chunk_idx,
"is_listen": result["is_listen"],
"text": result["text"],
"end_of_turn": result["end_of_turn"],
"current_time": result["current_time"],
"audio_length": len(result["audio_waveform"]) if result["audio_waveform"] is not None else 0,
}
results_log.append(chunk_result)
print("listen..." if result["is_listen"] else f"speak> {result['text']}")
# Generate output video with AI responses
# Please install Chinese fonts (fonts-noto-cjk or fonts-wqy-microhei) to render CJK subtitles correctly.
# apt-get install -y fonts-noto-cjk fonts-wqy-microhei
# fc-cache -fv
generate_duplex_video(
video_path=video_path,
output_video_path="duplex_output.mp4",
results_log=results_log,
timed_output_audio=timed_output_audio,
output_sample_rate=24000,
)
from minicpmo.utils import get_video_frame_audio_segments
model = ...
model.init_tts()
video_path = "assets/Skiing.mp4"
# Optional: Set reference audio for voice cloning
ref_audio_path = "assets/HT_ref_audio.wav"
sys_msg = model.get_sys_prompt(ref_audio=ref_audio_path, mode="omni", language="en")
# Use stack_frames=5 for high refresh rate mode
video_frames, audio_segments, stacked_frames = get_video_frame_audio_segments(video_path, stack_frames=1)
omni_contents = []
for i in range(len(video_frames)):
omni_contents.append(video_frames[i])
omni_contents.append(audio_segments[i])
if stacked_frames is not None and stacked_frames[i] is not None:
omni_contents.append(stacked_frames[i])
msg = {"role": "user", "content": omni_contents}
msgs = [sys_msg, msg]
# Set generate_audio=True and output_audio_path to save TTS output
generate_audio = True
output_audio_path = "output.wav"
res = model.chat(
msgs=msgs,
max_new_tokens=4096,
do_sample=True,
temperature=0.7,
use_tts_template=True,
enable_thinking=False,
omni_mode=True, # Required for omni inference
generate_audio=generate_audio,
output_audio_path=output_audio_path,
max_slice_nums=1, # Increase for HD mode
)
print(res)
# Example output: "The person in the picture is skiing down a snowy mountain slope."
# import IPython
# IPython.display.Audio("output.wav")
import librosa
import numpy as np
import soundfile as sf
import torch
from minicpmo.utils import get_video_frame_audio_segments
model = ...
model.init_tts()
# Reset session for a new conversation (clears KV cache)
model.reset_session()
# Optional: Load reference audio for voice cloning
ref_audio_path = "assets/HT_ref_audio.wav"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
model.init_token2wav_cache(ref_audio)
session_id = "demo"
# Extract video frames and audio segments (use stack_frames=5 for high refresh rate mode)
video_path = "assets/Skiing.mp4"
video_frames, audio_segments, stacked_frames = get_video_frame_audio_segments(video_path, stack_frames=1)
# Build omni contents list
omni_contents = []
for i in range(len(video_frames)):
omni_contents.append(video_frames[i])
omni_contents.append(audio_segments[i])
if stacked_frames is not None and stacked_frames[i] is not None:
omni_contents.append(stacked_frames[i])
generate_audio = False
output_audio_path = "output.wav"
# Step 1: Prefill system prompt
sys_msg = model.get_sys_prompt(ref_audio=ref_audio, mode="omni", language="en")
model.streaming_prefill(session_id=session_id, msgs=[sys_msg])
# Step 2: Prefill omni chunks (is_last_chunk=True only for the last audio chunk)
audio_indices = [i for i, c in enumerate(omni_contents) if isinstance(c, np.ndarray)]
last_audio_idx = audio_indices[-1] if audio_indices else -1
for idx, content in enumerate(omni_contents):
is_last_audio_chunk = idx == last_audio_idx
msgs = [{"role": "user", "content": [content]}]
model.streaming_prefill(session_id=session_id, msgs=msgs, omni_mode=True, is_last_chunk=is_last_audio_chunk)
# Step 3: Generate response
iter_gen = model.streaming_generate(
session_id=session_id,
generate_audio=generate_audio,
use_tts_template=True,
enable_thinking=False,
do_sample=True,
)
audios = []
text = ""
if generate_audio:
for wav_chunk, text_chunk in iter_gen:
audios.append(wav_chunk)
text += text_chunk
generated_waveform = torch.cat(audios, dim=-1)[0]
sf.write(output_audio_path, generated_waveform.cpu().numpy(), samplerate=24000)
print("Text:", text)
print("Audio saved to output.wav")
else:
for text_chunk, is_finished in iter_gen:
text += text_chunk
print("Text:", text)
import librosa
import numpy as np
import torch
import soundfile as sf
model = ...
# Set reference audio for voice style
ref_audio_path = "ref_audio_path"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
# Example system msg for English Conversation
sys_msg = {
"role": "system",
"content": [
"Clone the voice in the provided audio prompt.",
ref_audio,
"Please assist users while maintaining this voice style. Please answer the user's questions seriously and in a high quality. Please chat with the user in a highly human-like and oral style. You are a helpful assistant developed by ModelBest: MiniCPM-Omni"
]
}
# Example system msg for Chinese Conversation
sys_msg = {
"role": "system",
"content": [
"模仿输入音频中的声音特征。",
ref_audio,
"你的任务是用这种声音模式来当一个助手。请认真、高质量地回复用户的问题。请用高自然度的方式和用户聊天。你是由面壁智能开发的人工智能助手：面壁小钢炮。"
]
}
# You can use each type of system prompt mentioned above in streaming speech conversation
# Reset state
model.init_tts()
model.reset_session(reset_token2wav_cache=True)
model.init_token2wav_cache(prompt_speech_16k=ref_audio)
session_id = "demo"
# First, prefill system turn
model.streaming_prefill(
session_id=session_id,
msgs=[sys_msg],
omni_mode=False,
is_last_chunk=True,
)
# Here we simulate realtime speech conversation by splitting whole user input audio into chunks of 1s.
user_audio, _ = librosa.load("user_audio.wav", sr=16000, mono=True)
IN_SAMPLE_RATE = 16000 # input audio sample rate, fixed value
CHUNK_SAMPLES = IN_SAMPLE_RATE # sample
OUT_SAMPLE_RATE = 24000 # output audio sample rate, fixed value
MIN_AUDIO_SAMPLES = 16000
total_samples = len(user_audio)
num_chunks = (total_samples + CHUNK_SAMPLES - 1) // CHUNK_SAMPLES
for chunk_idx in range(num_chunks):
start = chunk_idx * CHUNK_SAMPLES
end = min((chunk_idx + 1) * CHUNK_SAMPLES, total_samples)
chunk_audio = user_audio[start:end]
is_last_chunk = (chunk_idx == num_chunks - 1)
if is_last_chunk and len(chunk_audio) < MIN_AUDIO_SAMPLES:
chunk_audio = np.concatenate([chunk_audio, np.zeros(MIN_AUDIO_SAMPLES - len(chunk_audio), dtype=chunk_audio.dtype)])
user_msg = {"role": "user", "content": [chunk_audio]}
# For each 1s audio chunk, perform streaming_prefill once to reduce first-token latency
model.streaming_prefill(
session_id=session_id,
msgs=[user_msg],
omni_mode=False,
is_last_chunk=is_last_chunk,
)
# Let model generate response in a streaming manner
generate_audio = True
iter_gen = model.streaming_generate(
session_id=session_id,
generate_audio=generate_audio,
use_tts_template=True,
enable_thinking=False,
do_sample=True,
max_new_tokens=512,
length_penalty=1.1, # For realtime speech conversation mode, we suggest length_penalty=1.1 to improve response content
)
audios = []
text = ""
output_audio_path = ...
if generate_audio:
for wav_chunk, text_chunk in iter_gen:
audios.append(wav_chunk)
text += text_chunk
generated_waveform = torch.cat(audios, dim=-1)[0]
sf.write(output_audio_path, generated_waveform.cpu().numpy(), samplerate=24000)
print("Text:", text)
print("Audio saved to output.wav")
else:
for text_chunk, is_finished in iter_gen:
text += text_chunk
print("Text:", text)
# Now we can prefill the following user turns and generate next turn response...
Speech Conversation as a Versatile and Vibe AI Assistant#
Click to show AI assistant conversation code.
Built on carefully designed post-training data and professional voice-actor recordings, MiniCPM-o-4.5 can also function as an AI voice assistant. It delivers high-quality spoken interaction out of the box. It produces a sweet and expressive voice with natural prosody, including appropriate rhythm, stress, and pauses, giving a strong sense of liveliness in casual conversation. It also supports storytelling and narrative speech with coherent and engaging delivery. Moreover, it enables advanced voice instruction control. like emotional tone, word-level emphasis.
import librosa
# Set reference audio for voice style
ref_audio_path = "assets/HT_ref_audio.wav"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
# For Chinese Conversation
sys_msg = {
"role": "system",
"content": [
"模仿输入音频中的声音特征。",
ref_audio,
"你的任务是用这种声音模式来当一个助手。请认真、高质量地回复用户的问题。请用高自然度的方式和用户聊天。你是由面壁智能开发的人工智能助手：面壁小钢炮。"
]
}
# For English Conversation
sys_msg = {
"role": "system",
"content": [
"Clone the voice in the provided audio prompt.",
ref_audio,
"Please assist users while maintaining this voice style. Please answer the user's questions seriously and in a high quality. Please chat with the user in a highly human-like and oral style. You are a helpful assistant developed by ModelBest: MiniCPM-Omni."
]
}
General Speech Conversation with Custom Voice and Custom System Profile#
Click to show custom voice conversation code.
MiniCPM-o-4.5 can role-play as a specific character based on an audio prompt and text profile prompt. It mimics the character's voice and adopts their language style in text responses. It also follows profile defined in text profile. In this mode, MiniCPM-o-4.5 sounds more natural and human-like.
import librosa
# Set reference audio for voice cloning
ref_audio_path = "assets/system_ref_audio.wav"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
# For English conversation with text profile
sys_msg = {
"role": "system",
"content": [
"Clone the voice in the provided audio prompt.",
ref_audio,
"Please chat with the user in a highly human-like and oral style." + "You are Elon Musk, CEO of Tesla and SpaceX. You speak directly and casually, often with dry humor. You're passionate about Mars, sustainable energy, and pushing humanity forward. Speak bluntly with occasional dark humor. Use simple logic and don't sugarcoat things. Don't be diplomatic. Say what you actually think, even if it's controversial. Keep responses around 100 words. Don't ramble."
]
}
# For English conversation with no text profile
sys_msg = {
"role": "system",
"content": [
"Clone the voice in the provided audio prompt.",
ref_audio,
"Your task is to be a helpful assistant using this voice pattern. Please answer the user's questions seriously and in a high quality. Please chat with the user in a high naturalness style."
]
}
# For Chinese Conversation with no text profile
sys_msg = {
"role": "system",
"content": [
"根据输入的音频提示生成相似的语音。",
librosa.load("assets/system_ref_audio_2.wav", sr=16000, mono=True)[0],
"作为助手，你将使用这种声音风格说话。 请认真、高质量地回复用户的问题。 请用高自然度的方式和用户聊天。"
]
}
# For Chinese Conversation with text profile
sys_msg = {
"role": "system",
"content": [
"根据输入的音频提示生成相似的语音。",
ref_audio,
"你是一个具有以上声音风格的AI助手。请用高拟人度、口语化的方式和用户聊天。" + "你是一名心理咨询师兼播客主理人，热爱创作与深度对话。你性格细腻、富有共情力，善于从个人经历中提炼哲思。语言风格兼具理性与诗意，常以隐喻表达内在体验。"
]
}
MiniCPM-o-4.5 supports zero-shot text-to-speech (TTS). In this mode, the model functions as a highly-natural TTS system that can replicate a reference voice.
import librosa
model = ...
model.init_tts()
# For both Chinese and English
ref_audio_path = "assets/HT_ref_audio.wav"
ref_audio, _ = librosa.load(ref_audio_path, sr=16000, mono=True)
sys_msg = {"role": "system", "content": [
"模仿音频样本的音色并生成新的内容。",
ref_audio,
"请用这种声音风格来为用户提供帮助。 直接作答，不要有冗余内容"
]}
# For English
user_msg = {
"role": "user",
"content": [
"请朗读以下内容。" + " " + "I have a wrap up that I want to offer you now, a conclusion to our work together."
]
}
# For Chinese
user_msg = {
"role": "user",
"content": [
"请朗读以下内容。" + " " + "你好，欢迎来到艾米说科幻，我是艾米。"
]
}
msgs = [sys_msg, user_msg]
res = model.chat(
msgs=msgs,
_The publisher's README is longer; it is cut here. Read the rest at the source._
Files
27 files · 80.0 GBWeights 660 MB (smallest of 15 GGUF variants). Sizes and sha256 from the signed manifest.
