Raon-1 mini preview

KRAFTON's 28.8B mixture of experts (4.3B active) trained from scratch, strong in Korean, with a 262K context and a 17.6 GB NVFP4 build.

In short. Raon-1 mini preview is a text model from KRAFTON, the South Korean game company, trained from scratch to chat well in Korean and to drive coding and tool-using agents. It has 28.8 billion parameters, but only 4.3 billion work on each token, so it runs faster than its size suggests. A 4-bit NVFP4 build is about 17.6 GB, and KRAFTON says it runs on one 32 GB RTX 5090. The license is Apache 2.0, and the card calls this a preview, so the final model may differ.

Summary

Raon-1 mini preview is a text model that KRAFTON AI, the AI team of the South Korean game company KRAFTON, published on Hugging Face on 21 October 2026. KRAFTON says it designed and trained the model from scratch instead of adapting another lab's weights. It is a mixture-of-experts model with 28.8B parameters, of which 4.3B are active for each token (6 of 128 routed experts plus one shared expert), and a native context of 262,144 tokens. The card calls it a preview and says specs and results may change before the final version. It was trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, and its stated strengths are Korean conversation and agentic tasks.

Attention is a hybrid. Of the 25 layers, 18 use Raon Delta Attention, a linear-attention design with a fixed-size state that separates the erase and write gates of Kimi Delta Attention, and only the 7 gated MLA layers keep a KV cache, about 8 KiB per token in BF16. By our arithmetic a full 262,144-token context needs about 2 GiB of cache in BF16. The model is built to think before it answers, and the card advises against switching thinking off. The main repository holds MXFP8 weights of about 32.7 GB. KRAFTON also publishes an NVFP4 build of about 17.6 GB that it says runs on one RTX 5090, GGUF files from 16.1 GB to 30.7 GB, an MLX build for Apple silicon, and a 2.7 GB DFlash2 draft model for speculative decoding, which it measures at up to 2.18 times the output throughput on one B300.

KRAFTON's table compares it with two other mixture-of-experts models, Qwen3.6-35B-A3B and Gemma 4 26B-A4B, and with two dense models for reference, Qwen3.8-27B and Muse Glimmer 30B. Against the two MoE rivals it is ahead on GPQA Diamond (84.3), AIME 2026 (93.3), IFBench (84.2), SWE-bench Verified (70.8, mini-SWE-agent harness), BrowseComp Plus (55.7) and its internal Korean chat test (92.9). It trails Qwen3.6-35B-A3B on MMLU-Pro (76.5 against 85.9), Terminal-Bench Lite (47.0 against 63.3) and BFCL v4 (59.8 against 64.7 and 65.8 for Gemma 4), and the dense Qwen3.8-27B scores higher on most of the remaining rows. The Korean test is an unreleased internal benchmark, and a GPT-family judge scored the runs.

The Apache 2.0 license covers the weights and the files in the repository, and each repository carries a LICENSE file. KRAFTON says it does not cover its training data, training methods, hosted API or playground, which have separate terms. The card lists recommended sampling settings (temperature 1.0, top-p 0.95, top-k 20, and a presence penalty of 1.5 outside coding and math) and setup notes for Hermes Agent, Pi and OpenCode. It runs in Transformers, in KRAFTON's beta fork of vLLM, in llama.cpp and in MLX.

Good for

  • Korean-language chat and agent work from a model that fits one 32 GB GPU in its 4-bit form
  • Coding and tool-calling agents in harnesses such as Hermes Agent, Pi and OpenCode, using the 262K context
  • Long documents on modest memory, since only 7 of the 25 layers keep a KV cache
  • Faster serving with the published DFlash2 draft model, if you can run KRAFTON's vLLM fork

Watch out for

  • It is a preview, and KRAFTON says specs and results may change before the final version
  • Every score is KRAFTON's own, the Korean chat result uses an unreleased internal test, and the model trails Qwen3.6-35B-A3B on Terminal-Bench Lite, BFCL v4 and MMLU-Pro
  • It is meant to run with thinking on, so replies include a reasoning trace, and the presence-penalty setting is applied by custom code in Transformers
  • The card's main serving recipe is a beta fork of vLLM, and Transformers needs trust_remote_code
  • It handles text only, and the 2.18 times speedup was measured on one B300 with the MXFP8 model

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the GGUF IQ4_XS variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Raon-1-mini-preview (MXFP8) FP8 28.8B 4.3B active 66 GB 33 GB 20 GB Gaming GPU
Raon-1-mini-preview-NVFP4 NVFP4 28.8B 4.3B active 66 GB 33 GB 19 GB Gaming GPU
GGUF IQ4_XS INT4 28.8B 4.3B active 66 GB 33 GB 18 GB Gaming GPU
GGUF Q4_K_M INT4 28.8B 4.3B active 66 GB 33 GB 22 GB Gaming GPU
GGUF Q8_0 INT8 28.8B 4.3B active 66 GB 33 GB 20 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Raon-1 mini preview?
Raon-1 mini preview is a text model from KRAFTON, the South Korean game company, trained from scratch to chat well in Korean and to drive coding and tool-using agents. It has 28.8 billion parameters, but only 4.3 billion work on each token, so it runs faster than its size suggests. A 4-bit NVFP4 build is about 17.6 GB, and KRAFTON says it runs on one 32 GB RTX 5090. The license is Apache 2.0, and the card calls this a preview, so the final model may differ.
Can I use Raon-1 mini preview commercially?
Yes. Raon-1 mini preview is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Raon-1 mini preview open source?
Raon-1 mini preview is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Raon-1 mini preview on my own computer?
Yes, on a good graphics card. The smallest variant, GGUF IQ4_XS, needs about 18 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Raon-1 mini preview released?
Raon-1 mini preview was released by KRAFTON on October 21, 2026.
How long a context window does Raon-1 mini preview have?
Raon-1 mini preview supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
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
Raon-1 nano KRAFTON 3.1B 256K Apache 2.0 (Permissive) 1.98 GB Laptop or phone