~/AI ML/krafton-releases-raon-1-nano-a-1-5b-bilingual-model-with-256k

KRAFTON Releases Raon-1 nano: A 1.5B Bilingual Model with 256k Context

KRAFTON AI has released Raon-1 nano, an open-source dense language model featuring 1.5B effective parameters, native support for a 256k token context length, and bilingual Korean and English capabilities. It is optimized for on-device deployment and supports tool calling across multiple formats like GGUF, MLX, and LiteRT-LM. This release brings long-context capabilities and advanced bilingual instruction-following to resource-constrained edge devices, making local AI applications more powerful for Korean and English speakers. The use of innovative pruning and attention mechanisms demonstrates how larger architectures can be efficiently scaled down for edge deployment. Raon-1 nano uses a hybrid architecture combining 18 Raon Delta Attention layers and 6 full attention layers, derived from a larger model through MoE-to-Dense pruning and knowledge distillation. It is distributed under the Apache 2.0 license with ready-to-use quantization formats for Apple Silicon, Android, and other local environments.

## BACKGROUND

Open-source language models are often adapted into smaller, efficient formats like GGUF or MLX so they can run locally on consumer hardware without relying on cloud APIs. Techniques like knowledge distillation and specialized context extensions enable compact models to handle massive document lengths and complex tasks efficiently.

## REFERENCES

## KEYWORDS

#AI/ML#Large Language Models#Open Source#On-Device AI#Multilingual

$ subscribe --daily