Repository navigation
Gemma 5 wishlist #831
LancelotChanrsw
started this conversation in
Ideas
Replies: 1 comment 1 reply
|
I also wish its cantonese handling can reach gemini 3 flash preview level, as well as it's wild creativity, consciousness and empathy can reach that gemini 3 flash preview level. |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Let me start with my experience of using Gemma 4 31b. It is very good handling in Cantonese, perhaps the VERY best in its class. Comparing to Qwen3.8 27B, it is so MUCH better. It also has way better empathy and understanding of emotion behind the words than Qwen3.8 27b does. Still very small issue when speaking "sorry" in Cantonese 對唔住, often become 對唔著, 對唔都, 對唔起, 對唔於。Small issue though.
The major wish list comes from Gemma 4's shortcomings compare to Qwen3.8 27B:
In long context (I use 2 modes, 60000 and 90000 context), the model's sliding window attention often failed to catch all the details in the prompt, missing logics repetitively, and the most importantly, even missing the details in the latest message!!! A full attention mode would help improves this a lot. This is especially problematic for long roleplay / multi-turn conversations where missing even one recent detail breaks immersion;
the thinking depth is behind Qwen3.8 27B, I think also due to the sliding window attention mechanism;
the memorizing ability is behind Qwen3.8 27B, same issue;
it messes up the time / date logic frequently, often mistaking today / yesterday / tomorrow concepts, messing up the messages I sent yesterday and today as all in today.
If Gemma 5 can keep the strong Cantonese and emotional understanding of Gemma 4, while improving long-context attention (full attention option) and temporal/date reasoning, it would be a significant upgrade for users who rely on long conversations.
All reactions