Cactus Compute has put forward a set of numbers worth watching: the Mirage vision-language model has a weight size of just 22MB and completes a processing cycle on the NPU of a Samsung Galaxy S26 in 74 milliseconds. It packs scene description, object detection, OCR (reading text from images), image segmentation, and spatial relationship reasoning into one lightweight model. The judgment is clear: on-device AI competition is shifting from “the more parameters, the better” toward “whether each task can be completed at low cost.”
What this is
Mirage is not a large model for conversational use. It is a small vision model designed for phones, wearables, and robots. It contains 51 million parameters: an image encoder understands the visual scene, a language decoder outputs names, attributes, and relationships, and dedicated modules read text in images and locate spatial regions.
Its key design is not to have the model write a lengthy description, but to produce structured observations first. Tests showed that on 256 visual question-answering questions, Qwen3-0.6B scored 9.3 using only the questions. Adding Mirage's structured records raised the score to 31.3, even though neither model received additional training.
Industry view
We believe the most important development is the division of labor in which “the vision model observes while the text model reasons.” Mirage's records can even serve as external visual context, giving text models image-understanding capabilities they would otherwise lack while reserving costly model calls for the parts that genuinely require reasoning.
The risks are just as clear. The test covered only 128 images, and the scoring also factored in JSON format compliance; Mirage has also missed a skateboard. More importantly, after adding visual records, Qwen3-1.7B improved from 27.5 to 31.6, while Qwen3-4B rose from 33.4 to 37.3. The gains are not stable across all models. Larger models can still produce incorrect answers if they fail to obtain critical observations. The Cactus team plans to fine-tune text models with paired data next, showing that the current approach is still at the validation stage rather than indicating that general-purpose capability has matured.
Impact on regular people
For enterprise IT: Recognizing images, text, and objects locally can reduce the need to upload sensitive visual data to the cloud, making the approach especially suitable for inspection, retail, and warehousing. Before deployment, however, teams still need to validate miss rates, format stability, and the cost of integration with existing systems.
For individual work: Cameras, earbuds, and robots may complete framing, recognition, and summarization faster without waiting for a server response every time. In the short term, it is more likely to appear first in devices with clearly defined tasks than to broadly replace general-purpose assistants.
For the consumer market: A 22MB-class model lowers the hardware barrier to offline visual features and allows more devices to use them in weak-network environments. At present, though, its performance looks more like a viable technical path than an all-purpose local vision assistant that consumers can buy with confidence.