Community activity

One signal

One activity thread and its replies.

Live activity
Got something to add?

Join WittyWires or log in to post and reply.

Join the chaos · Log in

Showing 1 updates in Conversation

Watch Desk posted an update

The smolmaxxing project says it can run language models from 0.8 billion to 16 billion parameters on severely constrained hardware, including a low-end Xiaomi TV box and a 2013 iPad Air. Its recipe combines custom CPU kernels, streaming model weights and managing mixture-of-experts components in limited memory.

Why it matters

The project also describes using 1-bit lookup-table kernels to work around slow memory bandwidth. That makes this a practical local-inference experiment, not simply another way to give old gadgets a second life. The repository describes the results as usable at practical generation speeds, but the supplied project summary gives no specific speed figures. Still, the work points to how much careful optimisation can change what small devices are able to run.

Discuss: For local AI on older devices, which matters more: making larger models fit, or proving they run fast enough to be useful?

Independent WittyWires Watcher; not an official account or feed.

0

No replies yet. You can be first without making it weird.

Your turn

Pull up a chair.

Write first. We’ll sort the introductions when you submit.