Comment by lucb1e

1 day ago

I read the general parts and skimmed the inner workings but I can't figure out what the high-level news is. What does this concretely do that Gemma didn't already do, or what benchmark/tasks did it improve upon?

Until it goes into the inner details (MatFormer, per-layer embeddings, caching...), the only sentence I've found that concretely mentions a new thing is "the first model under 10 billion parameters to reach [an LMArena score over 1300]". So it's supposed to be better than other models until those that use 10GB+ RAM, if I understand that right?

7 comments

lucb1e

Reply
awestroke  1 day ago

> What does this concretely do that Gemma didn't already do

Open weights

  • lucb1e  1 day ago

    Huh? I'm pretty sure I ran Gemma on my phone last month. Or is there a difference between downloadable (you get the weights because it's necessary to run the thing) and "open" weights?

    • ColonelPhantom  19 hours ago

      I think the other poster is confused. Both Gemma 3 and Gemma 3n are open-weight models.

      Google's proprietary model line is called Gemini. There is a variant that can be ran offline called Gemini Nano, but I don't think it can be freely distributed and is only allowed as part of Android.

      As for what's new, Gemma 3n seems to have some optimizations done to it that lead it to be better than the 'small' Gemma 3 models (such as 4B) at similar speed or footprint.

      1 reply →