AI BriefingCactus Compute Releases Mirage, a 22 MB Vision Model for Consumer Devices
·2026.11.07 04:46
Key point
Cactus Compute's Mirage model uses 51.3M parameters to perform multi-task visual analysis on mobile NPUs, enabling text-only LLMs to answer visual questions without fine-tuning.
Details
- Cactus Compute released Mirage, a 22 MB vision model designed for phones, wearables, and robots, featuring 51.3M parameters including a 7.8M image encoder and 40.8M language decoder.
- End-to-end latency on a Samsung Galaxy S26 is 74 ms on the NPU and 107 ms on the CPU, supporting scene descriptions, object detection, OCR, segmentation, and spatial relations in seven languages.
- Passing Mirage's structured output to a text-only LLM like Qwen3-0.6B improved VQA scores from 9.3 to 31.3 on VQAv2 without fine-tuning, though larger models showed mixed results due to format compliance issues.
- The architecture uses HConvGLU encoders and Simple Attention decoders, with a key finding that structured records alone often outperform adding reasoning traces for smaller text models.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
GPT 5.6 Sol: OpenAI's Best Vision Model to Date
GPT-5.6 Sol is OpenAI's Best Vision Model to DateMistral Releases OCR 4.1
Cactus Compute Releases Whistle, a 16.9 MB On-Device Speech Recognition Model