Raon-1 nano

krafton/raon-1-nano is KRAFTON AI's open-weight dense model: 1.5B effective parameters (3.1B total, including a 1.5B per-layer embedding table), a 262,144-token context, and Korean and English conversation with tool calling. It is designed for on-device use, with GGUF, MLX and LiteRT-LM bundles alongside the main checkpoint. No API host serves it yet, so it is listed as coming soon.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "KRAFTON/Raon-1-nano"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True)

Weights: Hugging Face · Paper: MoE-to-Dense. Watch the model page for the route.