Cactus · On-device vision model

Mirage

Cactus has released Mirage, a 51.3 million parameter vision model for phones, wearables, robots and other consumer devices. In a single pass it describes scenes, detects objects, reads text, generates segmentation masks, identifies spatial relations and fills structured forms, returning the result as JSON. End to end it takes 74 ms on a Samsung Galaxy S26 NPU or 107 ms on its CPU, and it fits in about 22 MB (a 21.6 MB file) using 2-bit Cactus Quants. Descriptions are available in English, German, French, Spanish, Italian, Dutch and Polish. A shared convolutional and attention encoder feeds detection, mask and OCR heads, and an eight-layer decoder writes the description and references to the detected boxes. Paired with Qwen3-0.6B on 256 VQAv2 validation questions, Mirage's structured record and trace gave 31.3 out of 100, against 22.2 with another image's record. Omitting the trace gave 33.1, and larger text-only readers showed no consistent gains. It installs with `pip install -U cactus-compute` and runs locally; the post gives no licence or star count. It was published on 6 November 2026.