Sign in
Pending benchmarksTracking since Nov 13, 2026

Is Raon-1 nano state of the art?Pending benchmarks.Announced Nov 11, 2026; no independent results yet.

KRAFTON AI's dense language model with 1.5B effective parameters (3.1B total, including a 1.5B per-layer embedding table). It mixes Raon Delta Attention (linear attention) with gated MLA layers, natively handles 262,144 tokens, and is trained for Korean and English instruction following, conversation, and tool use. Official pages say it was derived from Raon-1 mini preview (30B-A4B) by MoE-to-Dense pruning and knowledge distillation, then continued pre-training on 3T tokens, followed by supervised fine-tuning, on-policy distillation, and reinforcement learning. Weights are ungated on Hugging Face under Apache 2.0.

Announced facts

The details so far

Lab
KRAFTON AI
Announced
Nov 11, 2026
API
Not yet
Open weights
Published · Apache 2.0, ungated on Hugging Face at KRAFTON/Raon-1-nano, plus official GGUF, MLX, and LiteRT-LM builds (official model card and LICENSE)
Parameters
1.5B effective, 3.1B total including a 1.5B per-layer embedding table (Hugging Face checkpoint 3.2B with a 55M MTP module)
Architecture
Dense hybrid, 24 layers: 18 Raon Delta Attention (linear) and 6 gated MLA; 256-dimensional per-layer embeddings; Raon-mHC residual with 4 streams; tied LM head; no positional encoding (official model card)
Context
262,144 tokens (native)
Training
Derived from Raon-1 mini preview (30B-A4B) by MoE-to-Dense pruning and knowledge distillation, then continued pre-training on 3T tokens, then SFT, on-policy distillation from Raon-1 mini preview, and RL. Languages: Korean and English (official model card)
Runtimes
Transformers, KRAFTON's vLLM and llama.cpp branches, MLX, and LiteRT-LM (official model card)
Hosted access
No public self-serve API or lab playground that serves Raon-1 nano is stated. KRAFTON's Raon-1 mini site links to the Hugging Face card

Tags marked unsure are not stated on the lab's own pages.

Next steps

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)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: self-reported on KRAFTON's official model card, IFBench 75.9, MultiChallenge 40.5, FACTS Grounding 44.4, PersonaMem v2 (32k context) 38.8, and PrefEval 44.0, against the card's comparison set of Qwen3.5 2B, Gemma 4 E2B, MiniCPM5 2B, and LFM3 1.7B.

Lab claim: self-reported on KRAFTON's official model card, KMMLU-Redux 50.9 and CLIcK 65.2. Two further scores, Chatbot Korean 91.0 and Office Work Bench 49.2, are from internal benchmarks the card says are not publicly released.

Receipts

Sources

LABKRAFTON AI announcement on Xx.com/Krafton_AI/status/2120517124032692241 · Nov 11, 2026REPORaon-1-nano on Hugging Facehuggingface.co/KRAFTON/Raon-1-nano · Nov 11, 2026PAPERMoE-to-Dense method paper (arXiv:2605.28207)arxiv.org/abs/2605.28207LABKRAFTON AI homepagekrafton.ai/en/LABRaon-1 mini site (link to Raon-1 nano weights)raon-mini.krafton.ai/en
Editorial

Compare when it lands

Unscored: Raon-1 mini (preview). A scored comparison appears once both are on a shared board.

Tracking continues until an independent board lists it.