Cactus Compute Unveils Mirage Vision Model for Edge Computing
Cactus Compute has announced the release of Mirage, a small-scale visual-language model designed specifically for deployment on consumer hardware, including smartphones, wearables, and robotics. According to a technical announcement by Henry, a representative for Cactus Compute, the model is optimized to run locally on everyday devices rather than relying on cloud-based processing.
Technical Specifications and Performance
The Mirage model features a total of 51.3 million parameters. The architecture is divided into a 7.8 million parameter image encoder, a 40.8 million parameter language decoder, and 2.6 million parameters dedicated to OCR (Optical Character Recognition) and spatial heads. The encoder utilizes HConvGLU, which employs fixed Hadamard transforms and small learned convolutions to process visual data.
In performance testing conducted on a Samsung Galaxy S26, the model demonstrated an execution time of 74 ms on the Neural Processing Unit (NPU) and 107 ms on the CPU. The model’s weights file is approximately 22 MB. Cactus Compute stated that the model is capable of performing scene descriptions, object detection, segmentation, spatial relations analysis, and decision-making in a single pass. Language support for descriptions includes English, German, French, Spanish, Italian, Dutch, and Polish.
Benchmarking and Integration Potential
The company’s research suggests that Mirage can enhance the capabilities of existing text-only Large Language Models (LLMs) by providing structured visual data. In unverified testing reported on Reddit, the model was used to provide structured records to the Qwen3-0.6B model. On 256 VQAv2 questions across 128 images, the Qwen3-0.6B model’s score increased from 9.3 to 31.3 when provided with Mirage’s structured record and reasoning trace.
However, the effectiveness of this integration appears to vary depending on the scale of the text model used. While the Qwen3-4B-Instruct-2507 model saw an increase from 33.4 to 37.3, other models showed mixed results. The company noted that performance can be impacted by format compliance, as some models failed to adhere to the required JSON output format when processing Mirage’s data. Additionally, the company acknowledged instances of perception failures, where the model failed to identify specific objects in certain images.
Developer Reception
Initial feedback from the developer community has focused on the availability of open weights and the potential for offline utility. Users on Reddit have expressed interest in the availability of safetensors files and have inquired about quantization options to optimize the model for various device capacities. While some users have expressed interest in the model’s ability to function without an internet connection, others have noted difficulties in accessing specific features, such as decision-making prompts, within the current web application.