KRAFTON releases Raon-1 nano, a 3.1B parameter open-source model with 262K context

·2026.11.12 01:31

Key point

The model uses a hybrid dense architecture with 18 linear attention layers and supports Korean and English tool calling.

KRAFTON AI released Raon-1 nano, a dense language model with 3.1B total parameters (1.5B effective) under the Apache 2.0 license.
The model supports up to 262,144 tokens of context and is optimized for Korean and English instruction following and tool calling.
Architecture features a hybrid design with 18 Raon Delta Attention layers and 6 Gated MLA layers, using per-layer 256-dimensional embeddings.
Derived from Raon-1 mini preview via MoE-to-Dense pruning and distillation, with continued pre-training on 3T tokens.
Available in GGUF, MLX, and LiteRT-LM formats for local inference, including on Apple silicon.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.