Raon-1-mini-preview (Hugging Face model card)
B · Useful, limited KRAFTON AI's model card documents an open-weight Apache 2.0 release with full architecture and serving details, but all benchmark results are self-reported, partly model-judged, and the Korean result uses an unreleased internal test.
Summary based on the full text·Original: about 10 minutes to read
In short
The Hugging Face model card (created 21 October 2026) for Raon-1-mini-preview, a mixture-of-experts language model designed and trained from scratch by KRAFTON AI, the game publisher's AI unit, and released under Apache 2.0. Each token uses 4.3B of 28.8B parameters; 18 of 25 layers use a linear "Raon Delta Attention" with fixed-size state, so only 7 layers keep a KV cache. KRAFTON reports leading results among similar-sized MoE models on reasoning, instruction following, search and Korean conversation, and weaker results on knowledge and some agentic coding tests. It is labelled a preview whose specs may change.
Key numbers
- 28.8B / 4.3Btotal versus active parameters per token in the hybrid mixture-of-experts model
- 262,144tokens of native context length
- 70.8 vs 63.2SWE-Verified score against Qwen3.6-35B-A3B (mini-SWE-agent harness)
- 76.5 vs 85.9MMLU Pro score against Qwen3.6-35B-A3B, where Raon trails
- up to 2.18xoutput throughput gain from the DFlash2 speculative-decoding draft model on one NVIDIA B300
As reported by the source.
English·Web page·opens huggingface.co
What it reports
- Architecture: 128 routed experts, 6 active plus 1 shared; Raon Delta Attention modifies Kimi Delta Attention; Raon-mHC cuts the cost of DeepSeek's manifold-constrained hyper-connections; no positional encoding; 259,840-token vocabulary; about 8 KiB of KV cache per token in BF16.
- Against Qwen3.6-35B-A3B and Gemma 4 26B-A4B: GPQA Diamond 84.3 (vs 83.2, 80.7), IFBench 84.2 (vs 66.6, 80.9), SWE-Verified 70.8 (vs 63.2, 61.6), BrowseComp Plus 55.7 (vs 34.8, 32.2).
- Trails on MMLU Pro 76.5 (vs 85.9, 82.9), Terminal-Bench Lite 47.0 (vs 63.3, 52.0) and BFCL v4 59.8 (vs 64.7, 65.8). The dense Qwen3.8-27B leads on most tests.
- Korean chatbot benchmark: 92.9 against 74.5 for Qwen3.6-35B-A3B, on an internal, unreleased test.
- Trained on Korean, English, Japanese, Chinese, Spanish, French and Thai; tool calling with support for Hermes Agent, Pi and OpenCode.
- NVFP4 checkpoint runs on a single RTX 5090 (32 GB); DFlash2 draft model gives up to 2.18x throughput on a B300.
- Downloads at capture: 1,177.
How they got it
Developer-run evaluations; gpt-5.6-luna served as judge and user simulator for all evaluations. AutomationBench used the public split, not the official set; SWE-Verified used the mini-SWE-agent harness. No technical report or training-data details are in the card.
Limits
- All numbers are self-reported by the releasing lab, with comparison models chosen by it.
- Model-as-judge scoring and an internal Korean benchmark limit outside verification.
- Preview release; results may change. Training data and compute are not disclosed.
- Serving requires KRAFTON's vLLM fork and custom code (
trust_remote_code).
Why it matters for companies
A permissively licensed, mid-sized model that fits on one workstation GPU widens the options for teams that need on-premises deployment, especially for Korean-language work. Treat the benchmark table as a shortlist signal, not evidence: run your own tasks, and note the weak knowledge and terminal scores before using it for agentic coding. Custom serving code is an operational and security dependency to review before production.
Summary based on the full text. Read the original for the full argument and data.Read the original on huggingface.co