Nice work! :D
One of Honey-Yuzu's biggest strengths is its emotional intelligence, imo. To see that you've further improved it's EQ-bench score (to the point of rivaling the new kids on the block!) is really awesome! Looks like based DELLA linear is indeed the new meme.
Very interesting that you opted for Fimbulvetr-v2. I initially used v2 in a Honey-Yuzu prototype, felt like v2.1 was better in the merge, and I liked the improved context (16K-24K in testing). But I also for certain prefer v2 on its own over v2.1, and during some testing with Chaifighter-v3 I found v2.1 gave the whole model a first person bias. So I went back and tested my Honey-Yuzu prototype, and for some reason it just didn't do as well imo. So, I think your work here and choice of using DELLA were good moves and contributed to an overall better model.
I'm grabbing a GGUF, so I haven't tested it just yet. I'll do some side-by-side testing with Honey-Yuzu. In the meantime, you can find me on the Kobold and Backyard AI servers (matchaspeed). You should shoot me a DM or get ahold of me there!
Have a lovely day!!