2026-06-19 –, D105 (capacity 300)
AI output often has a distinctive and often disliked style. But does it have to?
Kimi K2 Instruct, an extremely large language model (1T parameter MoE), has developed a "vibe" uniquely attractive to technical users. with low sycophancy and enjoyable to read. It is not fully preserved in later Kimi models.
Can we transplant this "vibe", or at least style, to a smaller model? Can we use such a transplant to improve the style of technical text?
I chose the smaller IBM Granite 4 hybrid models as my distillation targets; the Mamba-hybrid architecture seems to take style transplantation well. The results so far are very imperfect, but interesting.
Principal Technical Writer at Red Hat in Ireland, with lots of side interests. Recently, AI has been an important part of those, and Misha had to navigate LLM fine-tuning with robots as the only mentors much of the time.