Megrez-3B-Omni

Megrez-3B-Omni is an on-device multimodal understanding LLM model developed by Infinigence AI (Infinigence AI). It is an extension of the Megrez-3B-Instruct model and supports analysis of image, text, and audio modalities. The model achieves state-of-the-art accuracy in all three domains: Image Understanding: By utilizing SigLip-400M for constructing image tokens, Megrez-3B-Omni outperforms models with more parameters such as LLaVA-NeXT-Yi-34B. It is one of the best image understanding models among multiple mainstream benchmarks, including MME, MMMU, and OCRBench. It demonstrates excellent performance in tasks such as scene understanding and OCR. Language Understanding: Megrez-3B-Omni retains text understanding capabilities without significant trade-offs. Compared to its single-modal counterpart (Megrez-3B-Instruct), the accuracy variation is less than 2%, maintaining state-of-the-art performance on benchmarks like C-EVAL, MMLU/MMLU Pro, and AlignBench. It also outperforms previous-generation models with 14B parameters. Speech Understanding: Equipped with the encoder head of Qwen2-Audio/whisper-large-v3, the model supports both Chinese and English speech input, multi-turn conversations, and voice-based questions about input images. It can directly respond to voice commands with text and achieved leading results across multiple benchmarks.