sidhant pfp
sidhant
@sidhant
open source is a hell of a drug. i dropped an ios port of gemma 3n yesterday, just a few days after google released it (with android-only support), and already someone else jumped in, added image understanding, polished the UX, and is prepping a PR. insane. this is what it’s all about — no NDAs, no permission slips, just vibes and shipping. gemma 3n is a frontier edge model, and rn we’re watching a legit community ecosystem form around it in real time. that’s magic. huge thanks to @TheMagicIsInTheHole on reddit — here’s the screenshot they shared 👇
6 replies
2 recasts
46 reactions

shoni.eth pfp
shoni.eth
@alexpaden
I’m still kind of baffled this can load into ram on a phone isn’t it like 4B param? I’ll have to read up but wow that image support is insane
1 reply
0 recast
1 reaction

sidhant pfp
sidhant
@sidhant
it's even more insane than that. "Gemma 3n models use selective parameter activation technology to reduce resource requirements. This technique allows the models to operate at an effective size of 2B and 4B parameters, which is lower than the total number of parameters they contain."
1 reply
0 recast
2 reactions

shoni.eth pfp
shoni.eth
@alexpaden
oh interesting I guess this must be a new form of moe? I’m out of the loop but outside of this I think another huge exciting step is gluing together high speciality/niche models (I want to train on social data somehow, just not sure how yet) Thanks for the comment + great work
0 reply
0 recast
0 reaction