KRAFTON open-sources Raon-1 mini: 30B MoE with 4B active params, strong coding and Korean scores
Kangwook_Lee · x · 2026-10-21
KRAFTON AI released a preview of Raon-1 mini, a from-scratch 30B MoE LLM with 4B active parameters, with Apache 2.0 open weights plus a DFlash drafter; the API is free for a limited time.
- Architecture: MoE + linear attention + speculative decoding for fast, memory-efficient inference
- Benchmarks: GPQA Diamond 84.3, IFBench 84.2, BrowseComp Plus 55.7, SWE-Verified 70.8 — leading similar-size MoE models on knowledge, science reasoning and instruction following (vs Qwen3.6 35B-A3B, Gemma 4 26B-A4B, Muse Glimmer 30B)
- Best AA-Omniscience hallucination score (3.5) among compared models; 92.9 on internal Korean colloquial eval
- Matches Muse Glimmer on coding/agentic tasks while being faster; bigger models and final release coming soon
Related event: KRAFTON Open-Sources Raon-1 mini, a From-Scratch 30B-A4B MoE Model(5 posts)→
More from Models
- SuperGrok in Cursor: user burns through weekly quota, 3-day wait for reset — mhdfaran · 2026-10-23
- AI uncovers PHP2Shell flaw threatening the tech behind 70% of all websites — AIFlow_ML · 2026-10-23
- OpenAI support wrongly cancels his workspace, wiping a 48-month promo worth ~$1,100 — Tile124 · 2026-10-23
- SemiAnalysis: Anthropic A/B testing ~20% lower usage limits on some accounts — imjustnewatai · 2026-10-23
- Swift-1.5-Qwen3.8-27B ported to hipfire MQ4: 2.2x faster end-to-end than GGUF on a 7900 XT — Queasy_Asparagus69 · 2026-10-23
- Boundless offers abliterated DeepSeek V4.1 Flash for defensive security testing — MoonL88537 · 2026-10-23