Has anyone tested gemma 4 local LLM in Android Stu...
# random
p
Has anyone tested gemma 4 local LLM in Android Studio? I'm doing a try, and seems to be very buggy
s
Yeah, local LLM mostly is not worth it. Minimax 2.5 on Macs with 128 GB is your best option.
p
128 gb ram? wow
I have 32gb
➕ 1
google announced Gemma 4 as the best option for Android
but when I ask to check my currect activity and show me ways to solve an issue, he ask that he needs to check my current activity. Then I tell him that he already has permissions to see all my context and then, he tells me: "oh yes I see your activity, now tell me what u want". Its a loop. He don't to anything more.
Just giving me some general tips, but not touches the code or obeis my orders
s
You can try different models (cloud and self hosted) through OpenCode
I use the free available MiniMax 2.5, which is said to require „only“ 128 GB VRAM. Fortunately Apple devices have unified memory, because you won't find a consumer graphics card with that much.
Gemma 4 failed my test.
Currently there is no local LLM model capable of coding that runs on my 16 GB NVIDIA VRAM.
p
I use LM Studio, supposedly it uses computer RAM and not graphics card RAM for runing local models
I have 32gb ram on this computer
sorry if I'm wrong with anything but I'm starting in local models understanding
I'm trying to avoid using any online models with Android Studio and project context, because I don't want to give my project source code to a third party.
a
Didn't they say that Gemma 4 was best option for running ON android? Not for Android Studio development.
p
If you need to use a local model, we recommend trying Gemma 4, a state-of-the-art local model available for agentic coding in Android Studio.
they say this
now I tryed with the huge model, Gemma 26B MoE
and the result was exactly the same
this is very frustrating
s
I wouldn't consider 26B „huge“. It's a medium model.
128B is where it’s getting useful, but that's still not the biggest size
265 Views