KRAFTON Open-Sources Raon-1 mini, a From-Scratch 30B-A4B MoE Model

The Raon team at KRAFTON (PUBG's parent company and a top Korean AI team) has joined the open-weights camp, releasing its first fully from-scratch trained open model, Raon-1 mini preview, on Hugging Face under Apache 2.0. It's the company's first self-developed open large model, and what makes it notable is that a gaming giant entered the open ecosystem via fully self-trained models rather than distillation or reworking—plus the engineering details (speculative decoding draft models, quantization down to consumer GPUs) are fairly complete.

Confirmed

• Model specs: a Mixture-of-Experts (MoE) language model with 30B total parameters and 4B active, trained completely from scratch, released as a preview on Hugging Face under Apache 2.0
• Architecture: 18 layers of linear attention (per @KangwookLee's post highlights)
• Built-in reasoning capability with standout performance on math and science benchmarks (per @jacek2023); official comparison charts show it leading multiple evals among same-size models while offering faster inference and lower memory usage
• Focused on Korean-language ability and agent tasks, with an API available
• Ships with DFlash / DFlash2 speculative decoding draft models; the main model runs on a single RTX 5090 after NVFP4 quantization

Unconfirmed

• All posts are secondhand accounts or official marketing; specific leaderboard numbers behind "beating same-size models on multiple benchmarks" and third-party replication results have yet to appear

Why it matters

• Marks a top Korean gaming company officially joining the open-weights camp, providing a new Apache 2.0 commercially usable option for Korean-language and agent use cases
• The combination of from-scratch training, lightweight active parameters, speculative decoding, and consumer-GPU deployment is practically appealing for local deployment users

2026-10-21 ~ 2026-10-22 · 5 related posts

• [source] KRAFTON open-weights Raon-1 mini, its first from-scratch model, under Apache 2.0 — Kangwook_Lee · 2026-10-21
• KRAFTON enters open-weight race with Raon-1 mini, its first from-scratch model under Apache 2.0 — Kangwook_Lee · 2026-10-21
• [source] KRAFTON releases Raon-1 mini, a from-scratch 30B-A4B MoE model under Apache 2.0 — jacek2023 · 2026-10-21
• KRAFTON open-sources Raon-1 mini: 30B MoE with 4B active params, strong coding and Korean scores — Kangwook_Lee · 2026-10-21
• [source] KRAFTON open-sources Raon-1 mini, a 30B-A4B MoE model running on one RTX 5090 — Kangwook_Lee · 2026-10-22