~/EDGE AI/mirage-a-22-mb-visual-language-model-for-edge-devices

Mirage: A 22 MB Visual-Language Model for Edge Devices

Cactus Compute has released Mirage, a 51.3M parameter, 22 MB visual-language model that performs scene description, object detection, OCR, segmentation, and spatial analysis in a single pass. The model achieves inference speeds of 74 ms on an NPU and 107 ms on a CPU of a Samsung Galaxy S26. This release demonstrates how complex multimodal intelligence can be efficiently compressed to run locally on everyday consumer hardware like mobile phones and wearables. It also shows that passing structured model outputs to text-only LLMs can grant them multimodal capabilities without requiring costly fine-tuning. Mirage consists of a 7.8M image encoder using HConvGLU, a 40.8M language decoder, and 2.6M parameters for OCR and spatial heads. The model supports multilingual descriptions and demonstrates that structured records can improve the VQA performance of base text-only models, though format compliance issues still exist.

## BACKGROUND

Visual-language models (VLMs) combine computer vision and natural language processing to understand and describe image content. Edge AI refers to running artificial intelligence models directly on local hardware devices rather than cloud servers, ensuring lower latency and better privacy.

## REFERENCES

## KEYWORDS

#Edge AI#Computer Vision#Visual-Language Models#Model Optimization#LocalLLaMA

$ subscribe --daily