Cactus Compute's Mirage Squeezes Six Vision Tasks Into a 22 MB Phone Model
Cactus Compute released Mirage , a 22 MB Apache-2.0 vision-language model with 51.3M parameters at 2-bit. Single forward pass returns scene description, object detection, OCR, segmentation masks, spatial relations and structured decisions.
Takeaways
- Cactus Compute released Mirage , a 22 MB Apache-2.0 vision-language model with 51.3M parameters at 2-bit.
- Single forward pass returns scene description, object detection, OCR, segmentation masks, spatial relations and structured decisions.
- Runs in 74 ms on Samsung Galaxy S26 NPU, 107 ms on CPU, fully on-device.
- Install with
pip install cactus-compute; callcactus.Mirage().describe(). - Weak on accented Latin text and non-Latin scripts: early testers reported 0/8 on accented words.
- Designed as a perception front-end that pairs with any text LLM without adapters or retraining.
Cactus Compute has released Mirage , a 51.3-million-parameter vision-language model that combines scene description, object detection, optical character recognition, segmentation, spatial analysis and structured decisions in one forward pass. The company reports that the roughly 22 MB model runs locally on a phone’s neural processing unit, or NPU.
The model card lists an Apache 2.0 license and more than 16,000 downloads during Mirage’s first month. Its compact size targets phones, wearables, robots and augmented-reality devices that lack the memory, power or connectivity for multi-gigabyte vision-language models.
One pass, six outputs
A single Mirage invocation returns a structured record containing descriptions, detected objects, geometry, recognized text and relationships between elements in an image. Developers can request the following output types:
| Output | Contents |
|---|---|
| Scene description | A compact summary of the image in English, German, French, Spanish, Italian, Dutch or Polish |
| Objects | Names and attributes generated from the image |
| Geometry | Bounding boxes and segmentation masks |
| Text | OCR results for visible writing |
| Relations | Spatial links between detected objects |
| Decisions | Label choices, bounded counts, short text and references to detected objects |
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
