I've observed something similar. In my experience, this isn't necessarily that the model is "thinking less," but that the reasoning-to-action alignment is weaker. The model may explicitly state a plan for example "I'll read the file first" but then skip that step and jump directly into editing. That usually shows up more in multi-step coding tasks where missing a single verification step can introduce subtle bugs.

If you're using the same inference parameters as the base model, I'd also look at whether the merged model's training/objectives prioritize faster action over exhaustive reasoning. Quantization (IQ4_XS) can also slightly affect long-chain reasoning consistency, though probably not enough to fully explain the behavior.

A few things worth testing:

1. Compare the FP16 or a higher-bit quantized version to rule out quantization effects.
2. Increase temperature slightly (or reduce it) and see whether reasoning consistency changes.
3. Explicitly instruct the model: "Always verify by reading files before editing. Never assume file contents." Sometimes these models respond well to workflow constraints.
4. Compare the generated reasoning length and task success rate across identical prompts instead of reasoning length alone.

I'd be interested to hear whether others have reproduced this with the same GGUF and llama.cpp setup, since it could be an interaction between the fine-tune, quantization, and inference backend rather than the model itself.

I would love to try out the higher bit versions, but on my hardware they are painfully slow because i only have 16gb of vram, and splitting the model into ram isn't really viable for dense models. I will try out different parameters and settings in llama.cpp, as well as changing my usual "instruction guidance".

Edit:

I tried with temp = 1.0 (which is the reccomended for general tasks) and it seems to be following its thoughts way more, but i don't like having temp this high because it makes it more unpredictable and too "creative", i will try with other temps and different prompting styles.

I have to say this, the model seems very capable even tho i always don't expect much from finetunes, but this is kind of different, it's reasoning is way more structured than the base and the instruction following is very good, also the tool calling is perfect for now which wasn't always the case with base. Very happy still.

using q6 on two v620pro (2*32gb vram) working flawless. but only started using like 4h ago. Will keep you posted

Quick note:
Strongly suggest Q4KM min for these applications; with Q6 / Q8 being recommended.
IQ4XS may be too low and/or work slightly differently; as the "mixture" for "IQ" quants differs from "Q" quants (llamacpp)

I'm planning on upgrading my rig to have enough VRAM to do it, but for now iq4_xxs is the best I can fit in 16gb. Any bit of offloading will result in massive drop in speed.

I'm also looking forward to create my own Q4 Imatrix quant from this, (more for educational purposes since this would be the first time for me), so if you could release the full precision it would be great!

Either way even with iq4_xxs is find the model really good still, as always good job to everyone!

For info full precision:

https://huggingface.co/DavidAU/Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-NEO-MAX-MTP-GGUF/discussions/3

You may also want to see "ablit" discussion thread as a re-done version / re-heretic version also exists too now.
I have not tested the new "re-heretic'ed" version as of this writing.