Mirage first look: Cactus 22 MB on-device vision model

By LLMTest Lab · Nov 7, 2026 · 4 min read visionopen-weightslocal-llm
On this page

On this page

  1. What it returns
  2. The only numbers that exist so far
  3. How it compares on paper
  4. Who should care

Not tested by us yet. Mirage ships as downloadable weights and a Python package that runs on your own device, with no hosted API we can call, and we have not run it ourselves. Everything below comes from Cactus Compute's launch post and the model card on Hugging Face. We'll update this post when we have compared its output against a general vision LLM on our own images.

Mirage is a 51.3 million parameter vision model from Cactus Compute that returns a scene description, objects, text, masks, relations and a decision from one pass over an image, in about 22 MB. The company reports 74 ms end to end on a Samsung Galaxy S26 NPU and 107 ms on its CPU. It is Apache-2.0 licensed and built for phones, wearables and robots, not for answering open-ended questions about a photo.

Mirage
Maker Cactus Compute
Announced 2026-11-06
Parameters 51.3M
Size About 22 MB (21.6 MB file, mostly 2-bit weights)
Latency (vendor-reported) 74 ms on Galaxy S26 NPU, 107 ms on its CPU
Languages for descriptions English, German, French, Spanish, Italian, Dutch, Polish
License Apache-2.0, per the model card
Access pip install -U cactus-compute, weights on Hugging Face

What it returns

The pitch is that one small model replaces a pipeline of separate ones. According to Cactus, a single call gives you:

  • Scene and objects. Names, boxes and attributes such as colour, shape or material. Names come from the language decoder, so they are not limited to a fixed detector class list.
  • Segmentation. A mask for each described object, as polygons or a bitmap.
  • Text. Writing in the image with a box and a detection score per region, read by a text reader inside the model.
  • Relations. Which object is attached to another, or which action is shown, linked back to the objects in the result.
  • Decisions. A choice from labels you supply, a count in a range, or a short text field, with probabilities for each option.

The output is structured JSON, so your app can follow a relation from a sign to its pole instead of parsing prose. That is a different job from a chat vision model, and it is why we file this one next to decision models and OCR tools rather than next to general assistants.

The only numbers that exist so far

Cactus does not publish a benchmark table against other vision models in its launch post. The one quantified experiment is about composition with a text model: pairing Mirage's structured record with Qwen3-0.6B on 256 VQAv2 validation questions from 128 images scored 31.3 out of 100 on VQA consensus credit. Giving the same model another image's record scored 22.2, and dropping the reasoning trace scored 33.1. Cactus itself says it found no reliable benefit from including the trace, and that follow-up runs with larger text-only readers did not show consistent gains.

That tells you the record carries real information about the image, but 31.3 is a low absolute score, and it measures a tiny text model reading Mirage's output, not Mirage answering questions. Nothing in the post compares Mirage's OCR or detection accuracy with a dedicated OCR engine or detector. Until someone does, the speed and size are the claims you can build a decision on, and the quality is unproven.

The post's own sample also shows the failure mode: the OCR misread the accent in "ARRÉT" on a stop sign, and the object detection scores were 0.64 for the stop sign and 0.56 for the pole. A reasoning trace is included for inspection, but Cactus notes it is generated text and can contain details the model inferred wrongly.

How it compares on paper

Mirage General vision LLM (cloud) Dedicated OCR engine
Output Structured JSON in one pass Free text or JSON on request Text and layout only
Where it runs On device, offline Provider's servers Usually on device or server
Open-ended questions Limited; choices and fields you define Yes No
Per-image fee None Per token Varies

For document-heavy work, our OCR and document parsing comparison is still the better starting point, because Mirage has no published document benchmark. For offline apps where text must stay on the device, the trade-off is the same one we cover in privacy-first LLMs for apps.

Who should care

  • Mobile and embedded developers who need captions, object locations and text in one call, and cannot ship a multi-gigabyte model or send photos to a server.
  • Robotics and wearable teams for whom a roughly 100 ms budget decides the architecture. Reproduce the latency on your own hardware first; the figures are Cactus's, measured on its own device.
  • Anyone who needs accurate OCR or open-ended image Q&A today. Wait for independent tests, or use a larger model.

How the 22 MB is reached matters for expectations: most quantised matrices use 2 bits, with embeddings and routing at 4 bits and some tensors left at 16-bit floating point. Aggressive quantisation usually costs accuracy, and Cactus does not publish a before-and-after comparison. For another small open model released the same week, see our Alpie Seed explainer.

Related articles

Alpie Seed explained: a 9B Apache coding model for laptops
169Pi's Alpie Seed is a 9B Apache 2.0 model built on Qwen3.5-9B that runs offline. Specs, vendor benchmark claims, hardware needs and who should care.
Best LLMs for privacy-first apps in 2026: HIPAA, ZDR, on-prem
Which LLMs meet HIPAA, SOC 2, and data-residency requirements in 2026? Four paths compared: ZDR API, SOC 2 cloud, EU data residency, and full self-hosting.