How those modes should work?
I briefly tested this model (my Q5 quants). Even on simple questions like "Tell me about the sea" and "2+2=?" reasoning was triggered. Can you tell a bit more how this should work?
This model is entirely built in a training-free manner, and given my current technical capabilities, I can only achieve this level for now. To be honest, I’m not satisfied with the model’s performance, but I’ve decided to release it publicly for two main reasons:
- To offer a reusable approach: I hope this provides an initial methodology for constructing hybrid models. Others can not only adopt this approach but also further refine it, thereby advancing the relevant technology.
- To serve as a low-cost starting point: For developers interested in building hybrid models, this model can act as a solid checkpoint. Continuing training from this checkpoint will significantly reduce the cost and resources required for further development.
You’re certainly welcome to use it directly, but as you may have already noticed, the model’s judgment of question difficulty is still not very accurate. I will continue iterating and improving this technique, and I warmly welcome any feedback or suggestions.
Of course, it’s more convenient to have a 2-in-1 model. I switched from instruct mode to thinking mode until I just stuck with thinking and stopped bothering to switch (lazy). Another improvement could be dynamic reasoning depth, similar to how oss models adjust, but it would be even better if the model could dynamically allocate tokens based on task complexity instead of wasting them.