The AI that looks at photos
How many photos
can it look at in a day?
Qwen3.8 27B quantised to 4 bits (Q4, UD-Q4_K_XL format, 17.55 GB) does not only read text: it sees. A photo does not go in as a photo, it goes in chopped into tokens —about 975 for a normal 1 MP photo, about 2,700 at high resolution (2.8 MP): one token per 32×32 pixels— and from there the model writes what it sees: a description, some tags or a condition report. Looking costs time; writing costs time too. Here you have both.
Try it
Pick a photo, pick what you want done with it.
—
Action
—
A sample answer, written beforehand. What plays at real speed is the timing: how long this card would take to look at the photo and write this text.
And without stopping…
One photo is seconds. A day is thousands.
—
What this is good for
Six jobs somebody does today by looking at photos one by one
—
Five machines
The same photo, five different machines
Photos looked at and described in a day (24 h straight, one request at a time) with the model and quantisation you picked above.
Where these numbers come from.
The reading and writing speeds of the RTX 3090 Ti are measured, on 21-09-2026 with llama.cpp: 1,158 tokens/s reading and 39.7 tokens/s writing, with Qwen3.8 27B quantised to 4 bits (UD-Q4_K_XL, 17.55 GB) and the KV cache at q8_0.
What it costs to look at a photo is measured too (21 Sep 2026): it takes one token per 32×32 pixels (974 tokens at 1 MP, 2,702 at 2.8 MP) and those tokens are read at about 750 a second, slower than text because they go through the vision encoder first. Describing a 1 MP photo: 1.2 s looking + 5.4 s writing. Every other card is estimated from its bandwidth and its compute.
The answers in the demo are written beforehand from what can actually be seen in each photo: they are realistic examples, not live generation. The formulas, one by one, are in the method.
The calculation engine (tw-engine.js) is not there yet. Reload the page in a moment.