I played around with Gemma4 26B A4B for a few days.
Gemma4 officially supports reasoning (also called thinking mode).
When you turn this on, the reliability of answers increases significantly.
However, it consumes speed and tokens.
My system isn't powerful enough to run a 26B model, so I have to allocate many layers to the CPU
and that's the only way Gemma4 26B model runs. That's why it's a bit slow. But when I turn on this reasoning
it becomes incredibly slow. Even simple answers require waiting 20 seconds before displaying on screen, and questions that require some thought take 40 to 50 seconds before
showing an answer. The answers themselves are smart. But the wait is too long. On top of that, since my hardware is lacking, I can't keep a long context either.
When I turn off reasoning, I don't see the need to use Gemma4 26B.
Gemma4 26B supports 256k context, but my system can't handle it, so it's just pie in the sky.
After thinking about it coldly, I concluded that I should only use this for occasionally complex questions or one-off inquiries.
Before this I was using the Qwen3.5 9B model, but the Gemma4-E4B model also seems usable, so I'm planning to just use Gemma4-E4B normally.
