Gemma > gemini?

Gemma
1

I never played around with previous versions of Gemma because they were weak, weak compared to Gemini, which was my main tool. Hence my surprise at version 4, especially GEMMA-4-31B-it!

  1. How on earth does this model follow system instructions so absurdly well in my tests?! At Opus level! Beating every Gemini model out there hands down.

  2. No more triggers that worked sometimes and sometimes didn’t in Gemini models. In Gemma, system instructions work ALWAYS! Meanwhile in Gemini, 30% of generations were flawed even at the text formatting level, wasting requests.

  3. No crazy and hyperactive default persona! This is my biggest surprise. It is simply a plain AI model. I don’t need to add 10 prohibitions to my prompt about how to behave, because Gemma just answers the question without hyperactive behavior like immediately editing files, deleting things, throwing itself into a frenzy of frantic action.

The only downside, especially for my use case, is the lack of a 1 million token context window. Unfortunately, when I use it for creative writing, 256k tokens is barely a short story for an afternoon.

I built my own Gemma-cli to mend the broken heart left by the shutdown of Gemini-cli.

image
image1114×628 22.4 KB

I have everything I need. System instructions, completely my own, context files I can load as many as I want, immediately loaded into memory, full control over model parameters. On top of that, tools, bulk web searching, etc. I have a feeling the people behind Gemma are true enthusiasts, because unfortunately Gemini has been letting me down for a year now.

2

Hi @Dawid_M -

Thanks for sharing your experience and detailed comparison. It’s great to hear how well gemma-4-31B-it is working for you, especially with instruction following. Nice to hear you built your own Gemma-cli around it as well.