Post #2655254
2026-04-21 00:55 UTC
What makes me happy is that local small models are now good enough to be smart auto-complete.
I hate agentic AI coding, it devours my soul. But just a smart autocomplete that tries to predict what I was about to type and just helps me to stay in flow is a good thing, and it can be done on a consumer GPU, demanding less watts to run than any AAA videogame.
Replies (1)
-
@marius851000@mastodon.mariusdavid.fr 2026-04-23 06:35
@inex@pony.social What model do you use for autocomplete? Cause I know good "small" model for agent (gpt-oss:20b) (and have been using occasionally over the past months), but as far as autocomplete goes, I put gemma3:1.5b and the result are pretty much unusable (apparently not targetting this use case. Might try putting the same gpt-oss:20b in it)