This model
Good model, my usage experience is impressive, especially since this model is 35Ba3b. However, I would like to point out that the model's occasional thinking is very short, resulting in poor quality of generated text. Also, its response to context is not very good; for example, one segment describes me observing pedestrians inside the car, while the next segment describes me climbing out from under the car – which is quite strange. Additionally, the censored mechanism is a bit strict, and there is no way to bypass it through prompt words. This might be a general issue with the qwen35 architecture.
Hmm, odd. I found this merge model quite far from censored, personally. That could just be my standard system prompt IDK. You can trigger thinking if you want it to always activate or at least much more often by adding the open think tag to the end of your prompt/just before it's reply. If you find the reasoning too short, have you tried telling it to think for longer about it's replies? It's fairly good at instruct, so that will probably help. On recast, I find the reasoning sometimes too long, and will always reason even though it doesn't add open think tags, so I think both should be possible to trigger with prompting.
In terms of context, continuity or sometimes not understanding the physical world, yes, I think that will be an issue for models, particularly of this this size and you might need to just sometimes give it an (OOC:) or [OOC:] instruction for the thing that it's dumbly missing. It's also possible that the LM head swap for the prose bump slightly interfered with it's smarts, even though I interpolated it, I was slightly less conservative than I could have been because I wanted the prose to be notably better than qwen base. You could also try recast for physical reasoning if you are in ST.
Worth mentioning that Qwen's default temperature preference is not very high. Typically well under 1 is the advice. I use variable temp with a max of 1.1. So it's possible you might also need to bump that down. Something like your temp/sampler settings could also very much interfere with it's ability to make sense of the scene.