RE: Qwen 3.8 27B Released - My observations

Your analysis is very interesting. I didn't know that dense models activate all parameters simultaneously, whereas MoE models only use a portion. That explains a lot regarding the difference in speed.

The jump in the score (8.838 vs. 6.862) is impressive, though I’m surprised it takes nearly three times longer to respond. Do you think that wait is worth it for the average user, or is it more for enthusiasts with high-end hardware?

I was also curious about them skipping version 3.7. Was there a technical reason for that, or was it just a coincidence?

One more question: with your setup (dual RTX 6000 Pro), how noticeable is the difference in temperature or power consumption when running this model compared to others?
Although my work focuses on travel and analysis in a different field, I’m also passionate about technology and everything it offers us today.

Thanks for taking the time to run these tests and share them. It’s clear you put a lot of hours into this.