Creating bounding boxes and OCR work both well. Creating an SVG of an image does not work well.

With Q4 everything below 150B should be fine. You can also run the -Flash variant of this model in Q1, but it is probably not usable.

It's a good model, way better than the models from half a year ago.

You posted about it here: https://sh.itjust.works/post/11722444 This image is different to your last image.

The images you posted looks like this in gray scale:

This map seems to be unrelated to the actual numbers that i found online, but i also have not found any good statistics for the whole world.

You can already get a taste of the future with this finetune: https://huggingface.co/TheDrummer/Rivermind-24B-v1

Here is a small snippet from their description:

Why Rivermind 24B v1? While other AIs struggle with basic tasks, Rivermind 24B v1 handles complex queries with the precision of a Dyson vacuum cleaning every last speck of dust. It’s not just an AI—it’s your future, optimized.

Ready to upgrade? Try Rivermind 24B v1 today and experience the difference—because tomorrow’s AI is here, and it’s powered by Intel’s cutting-edge processors. 🚀

There are not many models that support any-to-any, currently the best seems to be Qwen3-Omni, the audio quality is not great and it is not supported by llama.cpp: https://github.com/ggml-org/llama.cpp/issues/16186

Breakout 71 is a fun game.