Quality of this model is awesome (creative) ... but ...
I use this model for creating interactive fiction. And the quality is awesome. However it has the same problem that Qwen3.x have: overthinking. The reasoning of this model is gigantic. This leads to long wait times inbetween generating turns of the narrative.
@DavidAU have you perhaps considered using ThinkingCap as a base model, if possible? https://huggingface.co/bottlecapai/ThinkingCap-Qwen3.6-27B
I second this comment, this model is extremely slow compared to ThinkingCap. If OP can use as base model, we'll have a win-win.
Hey;
I hear you , and under normal fine tuning conditions I would agree.
However, the goals for this model included no major changes to the model's core aside from "IQ points".
When you modify reasoning in a model, this has major effects - some clear, some not.
We know from experience it is:
- easy to modify reasoning
- hard to stop it from damaging some parts of the model's functions.
Case in point:
This was an expansion, creative fine tune and reasoning was modified.
At 655 likes this model is powerful, but limited in scope.
It is very smart - for the use cases - but not as good for outside use cases.
Whereas FF711 is designed to be a smarter BASE model in all respects for all use cases.
It can be used as is and/or as a BASE model for primary tuning.
Reasoning was very mildly adjusted, most from the methods used to raise the IQ of the model across the board.
The reasoning/thinking block maybe about the same (size) for most use cases, however every thinking block is smarter, and as a result so it the output.
If you study the examples on the repo page you will see what I mean.
Compare against base Qwen 3.6 27B it is no contest.
RE: Base model - ThinkingCap.
The only base models we use are Heretics and/or straight from the arch source - IE Qwen.
Likewise for the training/merging pipeline.
Bottom line: Every time a model is trained, it can be damaged.
That being said, the source of FF711 can be used as a BASE model itself for ThinkingCap training.
Hey;
I hear you , and under normal fine tuning conditions I would agree.
However, the goals for this model included no major changes to the model's core aside from "IQ points".
When you modify reasoning in a model, this has major effects - some clear, some not.We know from experience it is:
- easy to modify reasoning
- hard to stop it from damaging some parts of the model's functions.
Case in point:
This was an expansion, creative fine tune and reasoning was modified.
At 655 likes this model is powerful, but limited in scope.
It is very smart - for the use cases - but not as good for outside use cases.Whereas FF711 is designed to be a smarter BASE model in all respects for all use cases.
It can be used as is and/or as a BASE model for primary tuning.Reasoning was very mildly adjusted, most from the methods used to raise the IQ of the model across the board.
The reasoning/thinking block maybe about the same (size) for most use cases, however every thinking block is smarter, and as a result so it the output.
If you study the examples on the repo page you will see what I mean.Compare against base Qwen 3.6 27B it is no contest.
RE: Base model - ThinkingCap.
The only base models we use are Heretics and/or straight from the arch source - IE Qwen.
Likewise for the training/merging pipeline.
Bottom line: Every time a model is trained, it can be damaged.That being said, the source of FF711 can be used as a BASE model itself for ThinkingCap training.
You are saying that ThinkingCap should train your model instead? I am getting better/shorter thinking and speed with ThinkingCap model, a combination of this two models would be great.