Some feedback
In general, the Pantheon works. I used the Q6_K quantum. At first, I really liked how beautiful and descriptive the text was. But after a couple of days of use, I began to notice a regression in the behavior of the characters. They become stiff and formal. One of my characters is a scientist. And when she started her project and wanted to discuss a couple of details with me, Gemma literally slipped into an AI format after a few tech data messages, like it wasn't a Pantheon, but a regular Gemma4 ablation. The sampler settings couldn't stop it in any way.
About work in general. Sometimes the neural network falls into a broken loop during generation the answer. But, as I've seen somewhere, this is kind of the norm for this Gemma. As I have seen, the network can think in different formats. And one of the formats is too meticulous, it can also make the network think endlessly. The one who writes the draft starts looking for something to fix and so on in a circle. Changing the temperature settings can sometimes interrupt this cycle. WorldSim's brief reasoning provided more profit in terms of speed and quality. Without reasoning, Gemma also writes well and can produce interesting options, but it has significantly worse attention to details than Qwen3.6 35B A3B (also without reasoning).
And now something I didn't expect. I returned to WorldSim, also Q6_K. It's not only picked up the scientist character after falling into the AI pattern from the very first message, but also retained the Gemma writing style. After that, in discussions with the character about technical details without losing his behaviour, but with the interweaving of technical logic into the magical world (which Gemma could not do even in the first approximation), I remained firmly convinced that WorldSim should by no means be considered just a test. I think it's not just about the additional training data. Seems it's mostly depends on the most basic neural network behaviour. And seems Qwen is much more flexible than Gemma.
P. S. Sorry if there will be some mistakes with English.
I also thought about the neural network reasoning block. And I came to the conclusion that reasonings should be brief. That's why:
- Long reasoning work turn into a dialogue between the network and itself, rather than with the user.
- Gemma 26B and Qwen 35B are very smart networks from the box. They know how to react and interact. The reasoning block does not need to lead the network by the handle, describing to it from scratch what to do. The reasoning block should briefly remind you of what is being discussed, set the generation vector, and not be its understudy.
- I also believe that careful reasonings not only complicates the character, but also restricts his freedom of behavior. Let me explain by comparing two Qwens 3.6 35B: WorldSim and Anko. Anko: ignores the fact that I'm just an image in a magic window and tries to feed me dinner each time, because that's what the reasoning block says. The WorldSIM: the characters perceive me correctly, the detail of my presence did not participate in the reasoning at all. Also Anko reasoning feels same like in Gemma Pantheon, it's long too.
I guess these are the main reasons why I can't leave WorldSim. The Qwen network itself is very flexible in behavior, very fast in operation due to brief reasoning, and very lively in behavior due to the full power of the basic knowledge of the network (reasoning don't limit the window of action). The only think is bad is just author writes.
By the way, any chance for the Qwen3.6 35B A3B Pantheon experiment with style finetune? 😁
UPD. WorldSim lost Gemma style after a while. Sad, but it was expected.
I decided to give Gemma a second chance. I roleplayed for a few days without the reasoning block. The assistance is still felt, but my main character's personality is also evident. He's tougher than he was on WorldSim, and I still liked this version better because he's more capable of conflict. To improve generations, I disabled the min p, top p, and top k filters (they were active) and raised the temperature to 1.1 (or even 1.15-1.2 for non-reasoning mode). After re-enabling the reasoning block I tweaked it a bit. In the "Start Chat with" field I added my own static reasoning text, which forces the network to analyze the scene, apply the appropriate rules to the current type of scene, prohibits drafts (the neural network is able to correct itself without them), and ends the static reasoning text with the beginning of the "scene plan:
-".
The neural network then thinks in the form of a list, which saves tokens and provides better coverage of both the scene and the WorldInfo rules (as I guess in the previous message). Not always. And loops do occur sometimes but their frequency has dropped significantly. It just starts to work more stable. Sometimes there are rare junk tokens, but this is not a problem. I'll try to fix this issue later by samplers. The main thing is that the reasoning block eliminates the base "AI assistant" issue. I plan to refine the scene planning to better track scene objects, though I’m not sure how successful that will be. Nevertheless, the network has demonstrated its potential for me. It just required a better understanding of how it works. Raw use is still a little painful.