Raon-1: KRAFTON's Korean-first 28.8B model and a 3.1B on-device sibling

KRAFTON AI released Raon-1 mini preview, a 28.8B mixture of experts trained from scratch, and the 3.1B Raon-1 nano distilled from it, both under Apache 2.0 with a 262K context.

  • New release
  • 2 min read
  • Published

KRAFTON AI, the AI team of the South Korean game company KRAFTON, has published two Apache 2.0 language models under the Raon name. Raon-1 mini preview appeared on Hugging Face on 21 October 2026, and Raon-1 nano followed in early November.

What is new

Raon-1 mini preview is a 28.8-billion-parameter mixture of experts that activates 4.3 billion parameters per token. KRAFTON says it trained the model from scratch for Korean conversation and agentic tasks, using Korean, English, Japanese, Chinese, Spanish, French and Thai. The card calls it a preview whose specs may still change. Only 7 of its 25 layers keep a KV cache, about 8 KiB per token, which makes the native 262,144-token context cheap to store.

In KRAFTON’s own table it beats Qwen3.6 35B-A3B and Gemma 4 26B-A4B on GPQA Diamond (84.3), AIME 2026 (93.3), IFBench (84.2) and SWE-bench Verified (70.8), and on an unreleased Korean chat test. It trails Qwen3.6 35B-A3B on MMLU-Pro and on Terminal-Bench Lite, where it scores 47.0 against 63.3.

Raon-1 nano is a dense model with 3.1 billion parameters, 1.5 billion of them in an embedding table, so KRAFTON calls it 1.5B-effective. It was pruned from the mini model and distilled from it, then trained further and finished with reinforcement learning. It targets instruction following, context following and tool calling in Korean and English. In KRAFTON’s table it leads Qwen3.5 2B, Gemma 4 E2B, MiniCPM5 2B and LFM3 1.7B on IFBench and the Korean tests, while trailing on function calling.

Running it

The main repository holds 32.7 GB of MXFP8 weights. KRAFTON also publishes a 17.6 GB NVFP4 build that it says runs on one 32 GB RTX 5090, GGUF files from 16.1 GB (IQ4_XS) to 30.7 GB (Q8_0), an MLX build and a 2.7 GB draft model for speculative decoding, which it measures at up to 2.18 times the throughput on one B300. The nano model is 7.2 GB in BF16, and its 4-bit GGUF, MLX and LiteRT-LM builds are about 2 GB. Transformers needs trust_remote_code, and the card’s main serving recipe is a beta fork of vLLM.

License

Each repository carries an Apache 2.0 LICENSE file covering the weights and files. KRAFTON says the grant does not cover its training data, methods, hosted API or playground, which have separate terms. The mini card also says access is granted on request, but the files downloaded without a login when we checked.

Why it matters

The four sovereign-project models, including K-EXAONE 2.0, all have 250 billion parameters or more, so none of them runs on one consumer GPU. Raon-1 mini preview is the first Korean language model in our catalog that fits a 32 GB card at 4-bit, and the nano model fits in about 2 GB. Every score is KRAFTON’s own and some benchmarks are internal, so test your own Korean tasks. See the Korean model list and small models for laptops.

Sources

Models in this story

Model Released Parameters Context License Smallest variant needs
Raon-1 mini preview KRAFTON 28.8B 4.3B active 256K Apache 2.0 (Permissive) 18 GB Gaming GPU
Raon-1 nano KRAFTON 3.1B 256K Apache 2.0 (Permissive) 1.98 GB Laptop or phone
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
K-EXAONE 2.0 LG AI Research 750B 37B active 256K Apache 2.0 (Permissive) 518 GB GPU server