A 22 MB weights file may sound like a technical footnote, but for Cactus Compute it is the starting point: Mirage, a new visual-language model, is designed to run locally on phones, wearables, and robots, not to be served from a cluster. Henry from Cactus Compute reported end-to-end measurements on a Samsung Galaxy S26: 74 ms on the NPU and 107 ms on the CPU for scene descriptions, object detection, OCR, segmentation, and spatial relations analysis in a single pass. Descriptions support English, German, French, Spanish, Italian, Dutch, and Polish.

The architectural choice matters more than the parameter count. Mirage has 51.3 million parameters: 7.8 million for the image encoder, 40.8 million for the language decoder, and 2.6 million in the OCR and spatial heads. The encoder uses HConvGLU, which mixes channels through fixed Hadamard transforms and small learned convolutions. Detection and masks share image features, while the decoder uses the Simple Attention and Hadamard MLP family behind Needle and Whistle. This is not extreme compression of a generalist model; it is an attempt to separate perception from prose generation. The model puts structure into the problem through spatial heads, a dedicated text reader, and short references that connect names to regions, while the decoder produces names, properties, relations, and a compact reasoning trace.

The most relevant result for people looking at local deployment is that Mirage's output can give a text-only LLM VLM-like abilities without fine-tuning. On 256 VQAv2 questions from 128 images, Qwen3-0.6B went from 9.3 with the question alone to 31.3 using Mirage's structured record and full trace, measured as VQA consensus credit out of 100. Neither model was fine-tuned. On another image's record the score was 22.2, while removing the trace increased it to 33.1. In that sample the image-specific observations helped, but the reasoning trace did not provide a reliable benefit: the structured response was enough.

Larger text-only models did not follow the same pattern. With the same prompt and Mirage observations, Qwen3-1.7B went from 27.5 to 31.6 and Qwen3-4B-Instruct-2507 from 33.4 to 37.3. But an operational detail clouds these numbers: the runs required a JSON object with a string-valued answer, and responses that broke that format received zero credit. Nemotron failed that requirement on 214 of 256 responses, and Qwen3-1.7B on 34. The scores therefore mix answer correctness with format compliance. This is not a secondary issue: in an on-device pipeline that connects a visual extractor to a local LLM, the ability to produce structured output reliably becomes a system requirement, not a prompting detail.

The structural limit also emerges from perception. In a photo with a skateboard, Mirage missed the skateboard and described cars, people, and shoes. Qwen guessed a skateboard from the question alone, then answered a car after reading Mirage's record and trace. A larger reader still needs the relevant observation to be present in the record. This brings back the central point: Mirage does not promise autonomous visual reasoning, but a compression of evidence that a text-only LLM can consume. If that compression loses the useful detail, no downstream model can recover it.

Cactus Compute announced that it will try fine-tuning the reader on Mirage records paired with questions and short answers. For those evaluating on-premise and edge deployments, this points in a specific direction: instead of bringing a heavy multimodal model onto the device, keep a small local extractor and let a self-hosted or local LLM consume structured records. The advantages are not only about latency, but also data sovereignty: the image can be processed entirely on the device and turned into text, reducing what leaves the network. The cost is a strong dependence on extraction quality and data format. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these trade-offs in enterprise contexts.