Mirage

A 51M-parameter open vision model in 22 MB: describes a photo, finds objects, reads text and fills forms on a phone.

In short. Mirage looks at a photo and returns a structured report: what the scene is, which objects are in it and where, what text is written, and how things relate. It is tiny, 22 MB, so it can run on a phone, a wearable or a robot with no cloud connection. It is not a chatbot and cannot answer free-form questions about an image on its own.

Summary

Mirage is a 51.3-million-parameter vision model from Cactus Compute, announced on 6 November 2026 under the Apache 2.0 license for phones, wearables, robots and other consumer devices. In one pass it describes a scene, finds objects, reads text, draws masks, notes relations between objects, and can fill in a form or pick from a list of choices that you supply. With Cactus's 2-bit quantization the model file is 21.6 MB, and Cactus reports 74 ms end to end on a Samsung Galaxy S26 NPU and 107 ms on its CPU.

It is not a chat model. The output is a structured record in JSON: a scene description, objects with names, boxes and attributes such as color or material, a mask for each object, text regions with boxes and a recognized string, and relations between objects. Object names come from a small language decoder, so they are not limited to a fixed detector class list. Descriptions are available in English, German, French, Spanish, Italian, Dutch and Polish. A request can ask for only some sections, and a form request returns a choice from labels you supply, a count within a range or a short text field. Inside, a convolutional stem and an attention encoder feed detection and mask heads, a separate text reader and an eight-layer decoder from the same Simple Attention family as Needle and Whistle. Most weights are quantized to 2 bits, while token embeddings stay at 4 bits.

You use it through the cactus-compute Python package or the native Cactus engine, which can load Needle, Whistle and Mirage side by side. Cactus also tested pairing it with a text-only model. Qwen3-0.6B, which never sees the pixels, scored 31.3 out of 100 on 256 VQAv2 validation questions when given Mirage's record and the question, against 22.2 when given the record of a different image, and follow-up runs with larger text models did not show consistent gains. The launch post and card give size and speed figures but no standard benchmark scores for detection, segmentation or text reading, and the post's own example shows an incorrect accent in recognized text, so test it on your own images.

Good for

  • Adding scene description, object finding and text reading to a phone, wearable or robot app with no cloud call
  • Turning a photo into structured data, or into a decision from labels you choose
  • Pairing with a small text model such as Needle 3, where Mirage supplies the perception and the other model does the reasoning
  • Browser demos, since the launch post runs it locally in a web page

Watch out for

  • It is a perception model, not a chat model, and it cannot answer arbitrary questions about an image by itself
  • No standard benchmark scores for detection, segmentation or text reading have been published in the launch post or card
  • It runs through the Cactus engine rather than common tools such as Transformers or llama.cpp
  • The result includes a generated reasoning trace that can be wrong, and Cactus's text-model test found no reliable benefit from including it
  • It was announced on 6 November 2026, so it has little track record

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Mirage (2-bit Cactus Quants, .cact file) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Mirage (2-bit Cactus Quants, .cact file) 51M 0.118 GB 0.059 GB 0.0354 GB Laptop or phone
Mirage (FP32 checkpoint) FP32 51M 0.118 GB 0.059 GB 0.0354 GB Laptop or phone

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Mirage?
Mirage looks at a photo and returns a structured report: what the scene is, which objects are in it and where, what text is written, and how things relate. It is tiny, 22 MB, so it can run on a phone, a wearable or a robot with no cloud connection. It is not a chatbot and cannot answer free-form questions about an image on its own.
Can I use Mirage commercially?
Yes. Mirage is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Mirage open source?
Mirage is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Mirage on my own computer?
Yes. The smallest variant, Mirage (2-bit Cactus Quants, .cact file), needs about 0.0354 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Mirage released?
Mirage was released by Cactus Compute on November 6, 2026.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Needle 3 Cactus Compute 121M 8K Apache 2.0 (Permissive) 0.0835 GB Laptop or phone
Whistle Cactus Compute 55M – Apache 2.0 (Permissive) 0.253 GB Laptop or phone
MiniCPM-V 4.7 OpenBMB 1.55B – 35.2B 3B active 256K Apache 2.0 (Permissive) 1.07 GB Laptop or phone
PaddleOCR-VL Baidu 960M – Apache 2.0 (Permissive) 0.662 GB Laptop or phone