AI

Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass.

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 07/11/2026, 11:47:26 UTC.

Briefings

Cactus Compute released Mirage, a 22 MB vision model running scene descriptions and object detection on consumer hardware. On a Samsung Galaxy S26, it achieved 74 ms on NPU and 107 ms on CPU. The model supports 7 languages and uses 51.3M parameters with specialized spatial and OCR heads. A key finding is that passing Mirage's structured output to a text-only LLM can provide VLM capabilities without fine-tuning, though results vary by model size and format compliance.

Key points

• Mirage achieves 74 ms inference on NPU and 107 ms on CPU for Samsung Galaxy S26.
• Mirage supports 7 languages including English, German, French, Spanish, Italian, Dutch, and Polish.

Context

This finding matters as it demonstrates how to compress intelligence to run locally on everyday consumer hardware while maintaining visual-language capabilities through structured reasoning traces.

Limitations

Results are limited to specific hardware and model configurations; larger text-only models showed mixed performance. Format compliance significantly impacts credit scores, with some models failing to meet requirements despite correct answers.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 07/11/2026, 11:48:42 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeicprzmpfq6fqyag6k34pdnc6teyxefy6wglke7n54e7sg7omhmrny