Mirage fits image understanding into 22 MB, and runs in 74 ms on a phone

Cactus Compute has published Mirage on Hugging Face, a vision model with just 51.3 million parameters that the company says takes about 22 MB and answers in 74 milliseconds on a Samsung Galaxy S26 NPU. The license is Apache 2.0.

The Hugging Face card describes Mirage as a model for image understanding and structured extraction. According to Cactus Compute’s launch blog, which is dated November 6, one small network produces scene descriptions, objects with attributes and boxes, segmentation masks, text read from images, relations between objects, and decisions or filled-in structured forms. The card lists seven languages: English, German, French, Spanish, Italian, Dutch and Polish.

A single 22 MB file that does many vision jobs

The size is the headline. Cactus says Mirage uses 2-bit quantization for most weight matrices, which brings the deployment file to 21.6 MB. Token embeddings and routing weights are kept at 4 bits, and some other tensors stay at 16-bit floating point.

On architecture, the blog describes a shared encoder that takes 512-pixel images and builds a 16 by 16 grid of 256 feature cells, followed by 16 convolution-based blocks and 12 attention blocks. Separate heads handle detection, masks and text reading, and an 8-layer decoder produces the output. Cactus also says each of its convolution layers stores about 8 times fewer weights than a full 512 by 512 matrix would.

Speed is reported for one device. Cactus puts end-to-end latency at 74 ms on the Galaxy S26 NPU and 107 ms on its CPU. The company names phones, wearables and robots as targets, and says Mirage runs in the same binaries as its earlier Needle and Whistle models.

Getting started is a Python call: after pip install -U cactus-compute, the card shows cactus.Mirage().describe("photo.png"), and the blog adds an extract() method for structured output. Native apps use cactus_load and cactus_describe. The weights ship as mirage.cact and mirage.safetensors, and the source code is on GitHub. Bounding boxes are returned on a 0 to 1000 scale relative to the image.

What a 51M model changes for app developers

The practical audience is developers building camera features that must work offline or without sending images to a server: reading text from a label, flagging objects in a feed, or turning a photo into a form. A model this small can sit inside an app bundle without a large download, and Apache 2.0 allows commercial use.

It also fits a pattern in recent open releases. Nvidia’s 117M-parameter Déjà View rebuilds 3D scenes from photos, but is limited to research use. Zyphra’s Zonos3-Mini speech model uses 100M active parameters and is free under Apache 2.0. Mirage is smaller than both, and its permissive license puts it in the second group, not the first.

For users, the effect would be features that run on the device itself, which matters for photos and documents that people prefer not to upload. That is a design argument made by the category, not a measured result in this release.

One narrow benchmark and several gaps

The evidence for quality is thin. The only benchmark in the blog is a test on 256 validation questions from VQAv2, covering 128 images, in which Mirage’s output was handed to a separate model, Qwen3-0.6B, to answer the question. That pairing scored 31.3 out of 100 on consensus credit. Using another image’s record scored 22.2, and giving the model no record at all scored 33.1.

That last number is higher than the Mirage-assisted score, so the test does not show that Mirage’s description helps this language model on VQAv2. It is a small sample and a test of a pipeline, not a head-to-head with other vision models, and Cactus compares Mirage with none. These are the company’s own figures and have not been independently verified.

Other details are missing. Neither the card nor the blog gives a context length. The 74 ms figure is for one phone, and no results are given for OCR accuracy, detection quality on standard benchmarks or segmentation. The card also lists no inference provider hosting Mirage, so trying it means running the Cactus package locally.

Early interest is visible on the card: it shows about 24,000 downloads last month and 300 likes, with two Spaces already built on it. What would settle the question is independent testing of detection, OCR and the claimed latency on devices other than the Galaxy S26.