The AI's are coming!

Software
1

A story of capability, told in pictures ! (all on this pc)

fossilcrew-eye2-b-20260803-183250
fossilcrew-eye2-b-20260803-1832501102×771 159 KB

fossilcrew-new-eye-20260803-183147
fossilcrew-new-eye-20260803-1831471761×635 105 KB

2

This incredible AI vision model is:

minicpm-v4.5:latest 0c40168f46d1 6.1 GB

3

  • llava:13b = 8.0GB, older architecture, CPU inference, could describe but struggled to read text

  • minicpm-v4.5 = 6.1GB, newer (Qwen3-8B + SigLIP2 vision encoder), same CPU budget
    model that’s genuinely better at vision (it reads screenshots, photos, watermarks, and identifies a Citroën DS21 correctly) while using 24% less memory.

  • smarter AND lighter.

  • Faster inference — fewer params to push through the CPU

  • More headroom — leaves RAM free for everything else the crew is doing

  • The vision encoder is better even though the model is smaller — MiniCPM’s SigLIP2 + Qwen3 base beats old LLaVA’s architecture at a fraction of the size

  • A 6.1GB model that reads a Citroën DS21 off a watermarked vintage photo is more capable than an 8GB model that can’t finish a sentence about a flowchart.

  • And the GGUF I downloaded for the actual eye agent (/llm/MiniCPM-V-4_5-Q4_K_M.gguf, 5.03GB + 1.1GB projector) is even leaner than the ollama package

  • so my production vision model is running an even smaller footprint than the 6.1GB I see in ‘ollama list’.