mlx-community

/

gemma-3-12b-it-4bit

like 4

Follow

MLX Community 4.4k

Here's what I see...

yes, something is broken
model needs an image in context window to work properly

Here's what I see...

ah, I misremembered, it's actually<pad>

I am using lm studio as well and the output is "<pad>" repeated again and again like the previous images. The max context size is also like 4k tokens instead of 128k I think.

Taking a look.

@pcuenq
Could you also confirm which base model was used: google/gemma-3-12b-pt or google/gemma-3-12b-it?
I am asking because the model card lists google/gemma-3-12b-it as the base model, but the link in the right menu points to google/gemma-3-12b-pt.

Can confirm this report on gemma-3-27b-it-4bit

For me too, it is indeed mlx-community/gemma-3-12b-it-4bit
Also, I can confirm that adding an image to context does make it work, here's a screenshot of the 12B model describing it... impressive.

Can confirm that gemma-3-4b-it-4bit have the same problem

Same issue here for gemma-3-12b-it-4-bit

I can confirm that the model now works properly with this commit.