"Miki" is a first try at distilling Kimi K2's style. I think it can sound fun for a 1B model - but also note ifeval is VERY low. The model learned to be expressive and forgot how to obey. It can also hallucinate just to stay fluent.

Enjoy with caution. Optimal temperature 0.6, like K2 itself - or maybe even 0.5. Can benefit from cranking repeat_penalty up to 1.3 and maybe even to 1.6

(A pebble is just a small piece of Granite.)

Downloads last month
53
Safetensors
Model size
1B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for ramendik/miki-pebble-20251115

Finetuned
(5)
this model
Quantizations
3 models

Dataset used to train ramendik/miki-pebble-20251115