Kvark Sparkle 27B – reasoning model trained to think with fewer tokens

We had problems with Qwen spending too much time reasoning about tasks instead of doing them. That's why we decided to fine-tune it to use fewer reasoning tokens without losing accuracy.

After a bunch of experiments we ended up with KVARK Sparkle 27B.

On our evaluations it uses 34–67% fewer reasoning tokens while getting +- the same scores as the original qwen.

Model is uploaded to Hugging Face with FP8 and GGUF versions available.

I'd like to see how it performs on other people's tasks. To see where it can do worse than qwen.