altbot v2.4 is here!!

just pushed a pretty significant update, swapped out the old gemma3 27b model for qwen3-vl:30b-a3b-thinking and holy shit the difference is noticeable

the cool thing about this model is it's very dynamic, simple images get quick, concise descriptions without overexplaining everything. and with complex images, it takes its time and actually thinks through what it's looking at. you get better alt-text without wasting compute on stuff that doesn't need it.

basically: smarter where it counts, efficient where it doesn't.

oh and here's the part that made me double-check my math: electricity usage dropped by MORE THAN 10x. complex images that used to cost ~0.950 Wh now only use ~0.092 Wh. this is thanks to MoE (mixture of experts) architecture, despite being a 30b parameter model, it only activates 3b worth of parameters at a time. so you get way better results while using a fraction of the power.

been testing this for a bit now and honestly really happy with the results. descriptions are more accurate, less hallucination-y, and the model describes way better than before.

as always, altbot is fully open source and processes everything locally, no data gets sent to third parties. your images stay private.

if you wanna support continued development:
☕ https://ko-fi.com/micr0byte
💜 https://github.com/sponsors/micr0-dev

source code: https://github.com/micr0-dev/Altbot

thanks for helping make the fediverse more accessible :3