Cactus releases Mirage: a 22MB on-device VLM doing detection, OCR and segmentation in 74ms
alexcovo_eth · x · 2026-11-07
Cactus Compute released Mirage, a tiny vision-language model for phones, wearables and robots:
- Only 51.3M parameters, 22MB with 2-bit Cactus Quants; a single pass handles scene description, object detection, OCR, segmentation masks, spatial relations and structured decisions
- Runs end-to-end in 74ms on Galaxy S26 NPU or 107ms on CPU, with a local browser sandbox
- Describes images in 7 languages (EN/DE/FR/ES/IT/NL/PL)
- Trades general language capacity to compete with SmolVLM2-500M, LFM2.5-450M and Florence2-230M
- Feeding its structured output to text-only Qwen3-0.6B lifted VQAv2 scores from 9.3 to 31.3/100 without fine-tuning
Related event: Cactus Compute Releases Mirage: A 22MB On-Device VLM(2 posts)→
More from Embodied
- Researcher achieves arbitrary code execution on Meta Ray-Ban smartglasses via OTA firmware extraction — evilsocket · 2026-11-07
- MACHINA Summit co-founder on Europe's Physical AI bottlenecks and VC gap — MarwaEldiwiny · 2026-11-07
- Wayve brings Wayve Labs to CoRL, booth 106, beyond-driving roadmap — Jamie_Shotton · 2026-11-07
- Claim: GPT-6 Astra's robotics skills aren't emergent — it trained on robotics-specific data — gleech · 2026-11-07
- Team builds SlashVibe, a physical controller for managing AI agent sessions — trepasito16 · 2026-11-07
- UCLA's SidewalkBench: A Sim Benchmark for Sidewalk Robot Navigation, CoRL Spotlight — zhoubolei · 2026-11-07