Original title: Cactus-Compute/mirage
Cactus Compute released Mirage 11 months ago on 6, officially stating it has 5130 million parameters and is approximately 22 MB after quantization to 2 bits using their proprietary method. Given an image, it returns structured JSON in one go: scene description, object names and bounding boxes, segmentation masks for each object, text in the image (with boxes and confidence scores), relationships between objects, and it can also perform multiple-choice judgments based on provided options or fill out simple forms; descriptions support English, German, French, Spanish, Italian, Dutch, and Polish. The model card itself is brief, with details sourced from the official blog.
Why It Matters Integrating description, detection, segmentation, OCR, and multiple-choice judgment into a single small model means phones, glasses, and robots no longer need to chain multiple models together
Upvote score 211 · Likes 319 · Downloads 28553
Reasons
- Officially claimed end-to-end latency of 74 milliseconds (NPU) or 107 milliseconds (CPU) on the Samsung Galaxy S26, with demos on the blog page running locally directly in the browser
- Object names are generated by a language decoder, not limited to a fixed detection category table; objects, attributes, and relationships in the output can be read as fields from a small scene graph
- Its JSON output can be passed to text-only large models to answer image questions: officially, Qwen3-0.6B scored 31.3 points on 256 VQAv2 questions, while records for other images were only 22.2 points
- Both weights and code are licensed under Apache-2.0
Usage
- Run pip install -U cactus-compute, then import cactus; cactus.Mirage().describe("photo.jpg"); the first call automatically downloads the engine and weights
- The describe function allows using sections=["objects", "text"], fields=["bbox", "score", "mask"], mask_format="polygons" to retrieve only necessary parts, and language="de" to change the description language; bounding box coordinates are relative to the image in the range 0–1000
- For judgment tasks: vision.extract("photo.jpg", {"sign": {"enum": ["stop", "yield", "speed limit"]}}) returns the selected option and probabilities for each option; native applications use cactus_load to load mirage.cact and cactus_describe to pass RGB pixels
Risks
- Speed and VQA scores are self-reported by the official team; they also state that adding reasoning traces does not consistently improve VQA performance (without traces it is 33.1), and switching to larger text models yields no consistent benefits
- In official examples, OCR misread accents on road signs, with detection scores ranging from 0.56–0.64; the reasoning process in the output is generated text and may contain inference errors
- Requires running with Cactus's proprietary engine and .cact format; it is not a transformers model and currently lacks ComfyUI support; this entry has not been run