Overview
Raon-1 mini is a preview model released on October 21, 2026 by KRAFTON AI, the AI arm of the South Korean game company behind PUBG, under Apache 2.0. It was designed and trained in house rather than fine-tuned from someone else's checkpoint.
The architecture reads like a survey of the 2026 playbook. It is a 30B-A4B sparse mixture of experts: 28.8B parameters in total with 4.3B active, six of 128 routed experts firing per token plus one shared expert. Attention is a 3:1 hybrid, with 18 linear attention layers alongside seven gated MLA layers. Only those seven keep a KV cache, at roughly 8KiB per token in BF16, and that is where the 262,144-token context comes from.
Both custom components stand on other people's shoulders. Raon Delta Attention separates the erase and write gates of Kimi Delta Attention, and Raon-mHC is a lower-cost take on DeepSeek's manifold-constrained hyper-connections. Add NoPE and a four-stream residual and you have most of the fashionable tricks of the year, recombined.
Quantisation options are generous: official MXFP8 and NVFP4 builds, with the NVFP4 version running inference on a single RTX 5090 at 32GB. A DFlash2 draft model ships alongside for speculative decoding, which KRAFTON says lifts output throughput up to 2.18 times on a single B300.
Key Features
- 30B-A4B sparse mixture of experts: 28.8B total parameters with 4.3B active. Six of 128 routed experts fire per token plus one shared expert, across 25 layers with hidden size 2,560.
- 18 linear attention layers plus 7 MLA layers: A 3:1 hybrid where only the seven gated MLA layers keep a KV cache, at roughly 8KiB per token in BF16. That is what supports 262,144 tokens of context.
- Trained from scratch, not fine-tuned: The model card states KRAFTON AI designed and trained it. Weights are Apache 2.0 and cleared for commercial use.
- Runs on a single RTX 5090: Official MXFP8 and NVFP4 quantisations, with the NVFP4 build running on 32GB of VRAM. Community estimates put a 4-bit build within reach of 24GB cards.
- DFlash2 draft model for speculative decoding: A draft model ships with the release. KRAFTON reports up to 2.18 times the output throughput of autoregressive decoding on a single B300.
- Seven languages, strongest in Korean: Training covers Korean, English, Japanese, Chinese, Spanish, French and Thai, with the best results in Korean. Tool calling supports Hermes Agent, Pi and OpenCode.
Use Cases
- Korean-first chat and customer support systems
- Long-context deployment where VRAM is tight and one card has to do
- Low-latency NPC dialogue inside games
- A tool-calling backend for agent workflows
Pros
- Apache 2.0 with no additional terms for commercial use
- 4.3B active parameters keep hardware requirements low enough for one card
- Linear attention holds the KV cache to about 8KiB per token
- Official quantised weights save you the compression work
- A speculative decoding draft model ships with the release
Pricing
Apache 2.0 licensed, with weights free to download on Hugging Face for commercial use. The official site offers a limited-time free API trial, and hosted access requires agreeing to separate Hosted Access terms. Self-hosting starts from a single RTX 5090 at 32GB.
Summary
A game studio training a 30B mixture of experts from scratch and open-sourcing it is more notable than any single benchmark line. KRAFTON's chief AI officer calls it a milestone after more than five years of research, with the full version and larger models still to come.
Remember what a preview means: the model card says specs and results may change, and every score is self-reported. Download counts and community feedback are early, so treat it as a signal rather than a drop-in replacement. The interesting part is where a company rooted in interactive entertainment will push the optimisation, and that sits around latency and cost, which makes it directly relevant to in-game AI.
Version History
- Raon-1 mini preview released (2026-10-21): KRAFTON AI publishes Raon-1 mini preview on Hugging Face under Apache 2.0: a 30B-A4B sparse MoE trained from scratch with 18 linear attention layers and 7 gated MLA layers, supporting 262,144 tokens of context. Official MXFP8 and NVFP4 builds ship, the latter running on a single RTX 5090, with a DFlash2 draft model for speculative decoding. A limited-time free API trial is available.