KRAFTON AI released Raon-1 nano, a small language model that follows Korean and English instructions and calls tools during conversations. Developers can download its weights, the learned internal numerical data, from Hugging Face. They can run it on devices through compatible runtimes or serve it through an API server.
The model card, a guide to the model, lists office work and mobile control as key use cases. It is an open model for development teams considering combined dialogue and task execution on devices, including Apple silicon.
1. A model targeting on-device execution and tool calling
Raon-1 nano has 1.5 billion effective parameters, the internal values used in actual operation. Its total is 3.1 billion, including layer-wise embedding tables with 1.5 billion parameters. The Hugging Face weights add a 55 million-parameter multi-token prediction module, which predicts multiple tokens together, bringing the package to 3.2 billion.
Under the Apache 2.0 license, the repository's weights and configuration files can be used, modified, and redistributed, including commercially. Development teams can choose how to deploy the open model in their own runtime environments.
2. A hybrid architecture for long contexts
The model supports up to 262,144 context tokens. Of its 24 layers, 18 use linear attention to carry information in a fixed-size state, while 6 use full attention with a key-value cache. Only some layers retain caches, rather than all layers. According to the company, the cache in BF16, a 16-bit floating-point format, is about 7KiB per token.
Training began by reducing the architecture of the larger Raon-1 mini preview model and transferring its learned results. This was followed by continued pretraining on 3 trillion tokens, further training an already trained model, and post-training.
3. Strengths in Korean dialogue evaluations
According to the company, its Korean dialogue score was the highest among the five models compared. The internal evaluation below averages counseling, customer service, fandom, and PUBG conversations.
Table: Model scores in the Korean dialogue evaluation
Conditions: Chatbot Korean, an internal performance benchmark. Scores average counseling, customer service, fandom, and PUBG dialogue. Raon-1 nano uses temperature 0.1, controlling response randomness, and presence_penalty 1.5. Comparison models use their model cards' recommended settings.
Unit: Score
| Model | Score |
|---|---|
| Raon-1 nano | 91.0 |
| Qwen3.5 2B | 42.2 |
| Gemma 4 E2B | 87.6 |
| MiniCPM5 2B | 41.6 |
| LFM3 1.7B | 58.3 |
Gemma is the closest comparison in this evaluation. These results are worth examining first for teams prioritizing everyday Korean conversation.
Using the general dialogue sampling settings above, the company reported instruction-following scores of 75.9 on IFBench and 40.5 on MultiChallenge.
For context evaluations with the same settings, the company reported 44.4 on FACTS Grounding and 38.8 on PersonaMem v2 with a 32k context.
With the same settings, the company reported 44.0 on PrefEval, averaging generation tasks over 10, 70, and 150 conversation turns. The model card also reported the highest scores among the comparison models on these five instruction and context tests.
Selection criteria differ by task. The company reported Raon's MMLU-Redux knowledge score as 81.1, behind MiniCPM5 2B at 84.0. Raon's temperature was 0.7 for this test.
At temperature 1.0, the company reported Raon's score of 56.3 on the BFCL v4 general function-calling evaluation, behind LFM3 1.7B at 67.2. Comparison models used their model cards' recommended settings.
With the same tool-use settings, the company reported NL2SH-ALFA terminal task scores of 58.0 for Raon and 62.6 for Gemma 4 E2B. This is a reason to examine results for intended tasks alongside Korean dialogue performance.
4. Runtime options and practical costs
The model card describes paths for generating responses directly or launching an API server. Below is the original command for starting terminal chat with a size-reduced GGUF model. It follows the procedure for first building KRAFTON's dedicated llama.cpp branch.
./build/bin/llama-cli \
-hf KRAFTON/Raon-1-nano-gguf:Q4_K_M \
--ctx-size 65536 \
--reasoning on \
--chat-template-kwargs '{"enable_thinking": true}'Teams can choose between on-device deployment and server hosting, but must invest in matching runtime formats to hardware. The vLLM example assumes one NVIDIA B300 and BF16 weights, while an MLX format is available for Apple silicon.
According to the card, RDA in the Transformers path processes prompt tokens sequentially, potentially slowing prefill, when the model processes long input contexts. The LiteRT-LM package uses 4-bit weights and fp16 activations, temporary numerical data created during computation. It limits context to 4,096 tokens and answers without a thinking process. GPU program compilation takes about 30–60 seconds on the first run.
Raon-1 nano stands out in Korean dialogue and instruction following in published comparisons, but selection should also weigh deployment conditions and task-specific evaluation results.
