Mirage
by Cactus-Compute
Cactus Compute's on-device vision model: 51.3M parameters in a 21.6 MB file that describes a photo, finds objects, reads text, draws masks and fills forms in one pass
Cactus-Compute/MirageDeploy Mirage
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
Mirage is a small vision model from Cactus Compute, released on 6 November 2026 under the Apache 2.0 license, built to run on phones, wearables, robots and other consumer devices. One call returns a structured record of an image: a scene description, named objects with boxes, attributes and segmentation masks, the text in the image with a box and score per region, relations between objects, and answers to choice or count questions you supply. It describes images in English, German, French, Spanish, Italian, Dutch and Polish. Cactus reports 74 ms end to end on a Samsung Galaxy S26 NPU and 107 ms on its CPU.
Architecture
51.3M parameters. A convolutional stem reads the image at 512 px into a 16 x 16 grid of 256 cells, followed by 16 gated-convolution blocks and 12 grouped-query attention blocks with four register tokens. Detection and mask heads read finer stem features, a CTC reader decodes the text crops, and an 8-layer decoder writes the description while referring to detected boxes. The released file uses 2-bit quantisation for most matrices, which is how the model fits in 21.6 MB, and runs on Cactus's own engine through the cactus-compute package.
Mixpeek SDK Integration
# Describe an image on the device (usage from the Cactus blog and model card).
# pip install -U cactus-compute # first use downloads the engine and weights
import cactus
vision = cactus.Mirage()
record = vision.describe(
"frame_0042.jpg",
sections=["objects", "text"],
fields=["bbox", "score", "mask"],
mask_format="polygons",
)
# Boxes use a 0 to 1000 scale relative to the image.
# Store the record with the file and timestamp it came from, next to an image embedding.Capabilities
- Scene descriptions in seven languages
- Open-vocabulary object detection with attributes such as colour, shape and material
- A segmentation mask for each described object, as polygons or a bitmap
- OCR with a box and detection score for each text region
- Relations between objects, forming a small scene graph
- Decisions and forms: pick from labels you supply, count within a range, or fill a short text field
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| Samsung Galaxy S26 | End-to-end latency, NPU | 74 ms | Cactus Compute blog, 6 November 2026; 107 ms on the CPU |
| VQAv2 (256 validation questions) | Consensus credit, Mirage record read by Qwen3-0.6B | 31.3 | Cactus Compute blog; 22.2 with another image's record, 33.1 without the reasoning trace |
Performance
We have not measured it. The figures are Cactus Compute's own. The VQAv2 number scores a small text model reading Mirage's output, so it measures the record as an interface for another model. Mirage runs on the Cactus engine, so it is not a model you import into a Transformers pipeline.
Common Pipeline Companions
Frequently Asked Questions
What is Cactus Mirage?
A 51.3M-parameter vision model from Cactus Compute, released on 6 November 2026 under Apache 2.0. In one pass it describes an image, detects objects with boxes and masks, reads text, lists relations between objects and answers choice or count questions, and its 2-bit release fits in 21.6 MB for phones and other devices.
Can Mirage run on a phone?
Yes, that is what it is built for. Cactus reports 74 ms end to end on a Samsung Galaxy S26 NPU and 107 ms on its CPU, and the browser sandbox on its blog runs the model locally. Native apps load mirage.cact with the Cactus engine; Python uses the cactus-compute package.
Does Mirage do OCR and object detection?
Both. It returns each text region with a box, the text read by a CTC reader and a detection score, and each object with a name, a box on a 0 to 1000 scale, attributes and a mask. Object names come from its language decoder, so they are not limited to a fixed class list.
What license is Mirage under?
Apache 2.0, as listed on the Hugging Face model card, which allows commercial use.
How do I search images Mirage has described?
Store each image's record (scene, object names, attributes, text) with the file and timestamp, and index it next to an image-text embedding from a model such as SigLIP 2. Filters and keyword search then run over Mirage's fields, and a description query runs over the embeddings.
Specification
Build a pipeline with Mirage
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free