Looks really promising, especially with hardware costs going through the roof from datacenter buildout.

1 / 4
2 / 4
3 / 4
4 / 4
Comments (4)

Want to leave a comment?

Sort by: Best
1
NecroSocial

Note this paper is not without it's detractors. You can find at least one of them arguing it out here: https://old.reddit.com/r/LocalLLaMA/comments/1vv6v00/freetokens_project_is_impressive/

reply permalink report gild save
1
Anonymous 83102ad7

Other than for privacy concerns, I don't see the benefit of the every day average user running models locally.

reply permalink report gild save
1
NecroSocial

Cost. Like for generative, something like Seedance 2.5 costs about one dollar per second of inference. So one unusable 30s generation is a waste of 30 dollars. Making a 5 minute video? $300 not counting failed generations. Make 70-80 minutes worth of frontier-quality video locally and you'll have saved money equivalent to the cost of a 5090.

Plus the privacy, plus not being reliant on companies run by wanna-be oligarchs.

reply permalink report gild save
[DELETED]