OPENBMB / END-TO-END OMNI / 2026

MiniCPM-o 4.5

TOTAL9.37Bparameters

ACTIVE9.37Bper token

DEPTH36 + encodersrepeated blocks

HIDDEN4,096main trunk

CONTEXT40Knative tokens

CHECKPOINT18.74GB4 shards · Apache-2.0

INTERACTIVE SYSTEM BLUEPRINT

MODEL MRI

39,960 INPUT + 1,000 OUTPUT · 1.26P arithmetic FLOPs · click any module for evidence

VERIFIED ADAPTER

INPUT40.0KIMAGE · VIDEO · AUDIO · TEXT ↔ SPEECH

TRANSFORMER STACK36L SHARED LANGUAGE TRUNKQwen-class dense GQA · vision/audio projectors · streaming TTS branch

shared pre-norm trunk · modality encoders and speech decoder remain modular

OUTPUT1.0KTOKENS

wraps every repeated block

DENSE GQA TRUNK · FULL-DUPLEX SCHEDULING模型可在继续接收音视频的同时流式生成语音，需要并行维护理解上下文和 TTS 状态。

Modality Encoders视觉 encoder 读取图像与视频，音频 encoder 提取语音特征，TTS decoder 将语言状态转换为流式声学 token。

TOTAL9.37Ball active

LLM36L4,096

VISION27Lpatch14

AUDIO24Lencoder

TTS20Lstreaming

CONTEXT40Ktokens