Raon-1 mini preview
Released Oct 21, 2026 · knowledge cutoff unpublished
- Status
- Unknown
- Location
- South Korea
- Type
- LLM
- Modality
- Text
- Context window
- 262K
- Max output
- —
- Speed
- —
- Price ($/MTok in / out)
- — / —
- Cost per task
- —
- Open weights
- Yes
Strengths & weaknesses
Strengths
- Built around natural, colloquial Korean: Krafton tuned dialogue, instruction following and response speed for Korean users first
- Hybrid linear-attention design keeps a KV cache on only seven layers, so long sessions stay light on memory
- Ships with a DFlash2 draft model for speculative decoding and an NVFP4 build that fits a single 32 GB RTX 5090
Weaknesses
- A preview, not the release: Krafton promises final weights and a technical report later, so behaviour may shift
- Broad-knowledge recall, some coding and long-context work are its soft spots, by Krafton's own table; it is a Korean-dialogue specialist more than a generalist
- Hugging Face weights are gated behind an access request, and there is no hosted API beyond a time-limited free playground
Krafton's first general-purpose LLM, released 2026-10-21 as a preview: 30B total / ~4.3B active MoE trained from scratch, mixing linear-style and full (MLA) attention, trained on Korean, English, Japanese, Chinese, Spanish, French and Thai. Benchmarks are Krafton's own model-card figures (GPQA Diamond 84.3, SWE-bench Verified 70.8 with the mini-SWE-agent harness, MMLU-Pro 76.5), judged with gpt-5.6-luna; the card's 'Terminal-Bench Lite' 47.0 is not Terminal-Bench 2.1, so terminalBench is null. Also reported: BrowseComp Plus 55.7, AA-LCR 67.3, and 92.9 on an internal Korean chatbot benchmark. Context is 262,144 tokens; card examples set max_tokens 32768, which is not stated as a cap. availability is self-host: weights are gated on request and no hosted API or OpenRouter listing was found, only a limited-time free playground.
Benchmarks3 of 10 reported
MMLU-Proretired
GPQA Diamond
SWE-bench Verified
SWE-bench Pro
Terminal-Bench 2.1
AIME (latest)retired
Humanity's Last Exam
LMArena Elo
GDPval-AA v2
ARC-AGI-2
For developers
API model strings
Not researched
Licence
Apache License 2.0 · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor