Is Raon-1 mini (preview) state of the art?Pending benchmarks.Announced Oct 21, 2026; no independent results yet.
KRAFTON AI's first general language model, a sparse mixture-of-experts model trained from scratch, released as a preview. It mixes Raon Delta Attention (linear attention) with gated MLA layers, natively handles 262,144 tokens and is tuned for Korean conversation and agentic tasks. Weights are on Hugging Face under Apache 2.0, behind an automatically approved access request.
The details so far
- Lab
- KRAFTON AI
- Announced
- Oct 21, 2026
- API
- Waitlist
- Open weights
- Published
- Parameters
- 28.8B total, 4.3B active (marketed as 30B-A4B)
- Status
- Preview; KRAFTON says specs and results may change before the final release
- Architecture
- MoE, 25 layers: 18 Raon Delta Attention (linear) and 7 gated MLA; 6 of 128 routed experts plus 1 shared active per token; no positional encoding
- Context
- 262,144 tokens (native)
- Training
- Trained from scratch by KRAFTON AI; languages: Korean, English, Japanese, Chinese, Spanish, French and Thai
- Weights
- Apache 2.0 on Hugging Face (gated, approval is automatic), in MXFP8 and NVFP4, plus GGUF, MLX and a DFlash2 speculative-decoding drafter
- Hosted access
- Free playground for a limited time; hosted API in private beta by email request, under separate terms (not Apache 2.0)
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: matches Muse Glimmer on coding and agentic tasks while being faster and more memory-efficient. The model card's own table is mixed against Muse Glimmer 30B (for example SWE-bench Verified 70.8 vs 72.0, Terminal-Bench Lite 47.0 vs 70.7, BrowseComp Plus 55.7 vs 52.8).
Lab claim: the DFlash2 drafter gives up to 2.18× higher output throughput than autoregressive decoding on a single NVIDIA B300.
Sources
LABKRAFTON AI announcement on Xx.com/Krafton_AI/status/2112772568776339709 · Oct 21, 2026REPORaon-1-mini-preview on Hugging Facehuggingface.co/KRAFTON/Raon-1-mini-preview · Oct 21, 2026LABRaon-1 mini playgroundraon-mini.krafton.ai/en · Oct 21, 2026Tracking continues until an independent board lists it.