Is KTransformer supports regular .safetensors model?

[vibe-Chen]

I do have download a full set of Deepseek-R1-Q4, and it works fine on my machine nicknamed "Fishbowl". There's a question however: the model is somewhat too big, not elegant enough for daily tasks.
So I wonder if I could load a distilled model from Deepseek like this, which is light enough even for my 6-year-old laptop to port with. And there's another question:

Is KTransformer supports regular .safetensors model?

for my knowledge, the answer is no - the arguments not includes one to specify you model path, only GGUFs instead. So any ideas?

[cyhasuka]

if I could load a distilled model?

Technically, yes, but there isn't any need for it. The distillation model u listed is a Dense structure, and using KT does not allow for any degree of acceleration.You can try the qwen3 family of MoE small-scale models, or simply switch to a different inference framework such as llama.cpp, vllm, etc.

Is KTransformer supports regular .safetensors model?

No.

[vibe-Chen]

Thanks, and I've found that there ARE GGUF files for those distilled models, also created by unsloth. Though that makes no sense for now