KRAFTON AI's first Raon language model, after the Raon-Speech, Raon-OpenTTS and vision-encoder releases. It is a 28.8B total / 4.3B active MoE (marketed as 30B-A4B) that the card says was "designed and trained from scratch by KRAFTON AI". The architecture backs that up: it matches no existing open model. The config declares its own Raon1MiniForCausalLM with a 259,840-token vocabulary, and the design is a set of KRAFTON variants on recent Chinese-lab ideas. 18 of 25 layers are Raon Delta Attention, which "decouples the erase and write gates of Kimi Delta Attention". The other 7 are gated MLA, so only those keep a KV cache (about 8 KiB per token in BF16). There is no positional encoding at all (NoPE). Raon-mHC (4 streams) is a cheaper version of DeepSeek's manifold-constrained hyper-connections. The MoE layers route 6 of 128 experts plus one shared expert with sigmoid routing, and use a new SiTU-GLU activation. Native context is 262,144 tokens. Trained on Korean, English, Japanese, Chinese, Spanish, French and Thai. Training tokens, data and compute are not disclosed.

This is explicitly a preview, "not a final model", with the final release and a technical report still to come. It is designed for thinking mode only. Weights are Apache 2.0 behind an auto-approved access gate, in BF16, MXFP8 and NVFP4 (the NVFP4 build fits a single 32 GB RTX 5090) plus GGUF and MLX. A DFlash2 speculative-decoding draft model is included, claimed at up to 2.18× output throughput on a B300.

The self-reported table compares against Qwen3.6-35B-A3B and Gemma 4 26B-A4B, all judged by gpt-5.6-luna, and the result is mixed rather than a clean sweep. Raon leads on IFBench 84.2 (66.6 / 80.9), SWE-bench Verified 70.8 with mini-SWE-agent (63.2 / 61.6), BrowseComp-Plus 55.7 (34.8 / 32.2), GPQA Diamond 84.3 and AIME 2026 93.3. It trails clearly on MMLU-Pro 76.5 (85.9 / 82.9), Terminal-Bench Lite 47.0 (63.3 / 52.0) and BFCL v4 59.8, which is typical of a 4B-active model trained without a large external base. KRAFTON's own Korean chatbot benchmark (92.9) is internal and unreleased. Not on Artificial Analysis at filing.

Model Details

Architecture MOE
Parameters 28.8B
Active params 4.3B
Experts 128 (top-6)
Context window 262,144
License Apache 2.0

Benchmark Scores

Benchmark Score Mode
MMLU-Pro 76.5 —
GPQA Diamond 84.3 —
AIME 2026 93.3 —
HMMT Feb 2026 84.1 —
IFBench 84.2 —
LiveCodeBench v6 81.6 —
SWE-bench Verified 70.8 mini-SWE-agent
Terminal-Bench Lite 47.0 —
BrowseComp-Plus 55.7 —
AA-LCR 67.3 —

Variants

Name Parameters Notes
Raon-1-mini-preview — 28.8B total / 4.3B active; 25 layers (18 RDA + 7 gated MLA), hidden 2,560; BF16, MXFP8, NVFP4, GGUF, MLX; DFlash2 draft model
moeopen-weightreasoningmultilingualefficiency

Related