Apache-2.0llmRuns locally28.8B
Raon-1-mini: Specs, License & How to Run
Apache-2.0 preview compact MoE (28.8B/4.3B) from KRAFTON AI, trained from scratch.
- License
- Apache-2.0
- Parameters
- 28.8B (active 4.3B)
- HF likes
- 148 (2026-10-24)
- First seen
- 2026-10-21
TL;DR. Raon-1-mini is a llm model from KRAFTON AI. Parameters: 28.8B (active 4.3B).License: Apache-2.0. It is listed under ChatGPT, Closed LLM APIs.
| Developer | KRAFTON AI |
|---|---|
| Type | llm |
| Architecture | Hybrid MoE (card model_type raon1_mini). 25 layers (18 Raon Delta Attention + 7 gated MLA); 128 routed experts, 6 selected per token, plus 1 shared expert. Marketing name 30B-A4B. HF safetensors.parameters sum to 28,783,218,402. |
| Parameters | 28.8B · active 4.3B |
| Modalities | text |
| Languages | ko, en, ja, zh, es, fr, th |
| Context | 262,144 tokens |
| License | Apache-2.0 — LICENSE on the Hugging Face card is Apache License 2.0. README: you may download, use, modify, and redistribute the model and weights, including for commercial purposes, subject to that license. Apache does not cover KRAFTON training data, methods, hosted API, or playground — those have separate terms. The card shows an extra_gated access form; API gated field was false on 2026-10-24. Not legal advice. |
| Hardware | 1× RTX 5090 32 GB (NVFP4, vendor) — BF16/MXFP8 card usedStorage is 32,714,641,804 bytes (~30.5 GiB). Card says NVFP4 enables inference on a single NVIDIA RTX 5090 (32 GB VRAM). That is vendor. Peak VRAM is not independently measured. (not independently tested) |
| Hugging Face likes | 148 (as of 2026-10-24) |
| Model card | https://huggingface.co/KRAFTON/Raon-1-mini-preview |
| First seen | 2026-10-21 |
Raon-1 mini (card name Raon-1 mini preview) is a compact Mixture-of-Experts language model trained from scratch by KRAFTON AI. Official hosts we opened on 2026-10-24: the card KRAFTON/Raon-1-mini-preview (created 2026-10-21T00:14:48Z; 148 likes / 548 downloads), siblings GGUF (44 likes / 7,661 downloads), NVFP4 (22 likes / 190 downloads), MLX (24 likes / 269 downloads), and DFlash2 (19 likes / 145 downloads), the Apache 2.0 LICENSE, and krafton-ai/Raon-1-mini-preview (3 GitHub stars, created 2026-10-21).
This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Another compact Apache-2.0 local MoE on this site: Flint (39.8B/2.5B coding). Datacenter multimodal: MiniMax M3.1 Flash.
The README calls this a preview, not a final model. Hugging Face pipeline_tag is text-generation. Text only.
What we verified (2026-10-24)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/KRAFTON/Raon-1-mini-preview | HF API |
API gated |
false (extra_gated request form on the card) | HF API + card |
| Created / last modified | 2026-10-21T00:14:48Z / 2026-10-22T22:30:01Z | HF API |
| Likes / downloads | 148 / 548 | same |
| License tag + file | apache-2.0 / Apache License 2.0 | HF cardData + LICENSE |
| Params (table) | 28.8B total / 4.3B activated | card |
| Marketing name | 30B-A4B | card |
| safetensors.parameters | 28,783,218,402 | HF API |
| usedStorage | 32,714,641,804 bytes (~30.5 GiB) | same |
| Experts | 128 routed; 6 + 1 shared active | card |
| Context | 262,144 tokens | card |
| Layers | 25 (18 RDA + 7 gated MLA) | card |
| GitHub | krafton-ai/Raon-1-mini-preview · 3 stars · Apache-2.0 | GitHub API |
How to run (from the card, not tested)
README: request access once on the card (it says approval is automatic), then:
pip install -U huggingface_hub
hf auth loginDocumented runtimes: Transformers (transformers==5.16.1), vLLM, llama.cpp, MLX. Thinking mode is the designed path; the card says disabling it is not recommended. Shared sampling (copied): temperature 1.0, top_p 0.95, top_k 20. Presence penalty 0 for coding/math, 1.5 otherwise.
The DFlash2 draft sibling is for speculative decoding. Card throughput rows on a single NVIDIA B300 stay vendor.
Hosted playground / API terms are not Apache-2.0. Read raon-mini.krafton.ai/en/terms.
License
LICENSE is Apache License 2.0. Card tag apache-2.0. Same tag on the GGUF, NVFP4, MLX, and DFlash2 siblings. README: Apache covers the downloadable weights and repo components; it does not cover training datasets, methods, hosted API, or playground. Not legal advice.
What we did not invent
- RTX 5090 / B300 tokens/s stay vendor.
- Card bench rows (including the internal Korean chatbot score) stay self-reported.
- We do not rank Raon against ChatGPT, Flint, or MiniMax M3.1 Flash.
Related: Flint, Flint vs Ornith-1.5-122B, MiniMax M3.1 Flash, DeepSeek-V4.1-Flash, vLLM. Pixel editors are a different search: PhotoCraft vs GIMP.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| GPQA Diamond (card) | 84.3 (self-reported) | https://huggingface.co/KRAFTON/Raon-1-mini-preview |
| SWE-Verified (card, mini-SWE-agent) | 70.8 (self-reported) | https://huggingface.co/KRAFTON/Raon-1-mini-preview |
| IFBench (card) | 84.2 (self-reported) | https://huggingface.co/KRAFTON/Raon-1-mini-preview |
Deploy options
- local: Hugging Face weights (MXFP8)
- local: Hugging Face GGUF
- local: Hugging Face NVFP4
- hosted: Raon playground (separate terms)
FAQ
Is Raon-1-mini Apache-2.0 and downloadable?
The card KRAFTON/Raon-1-mini-preview (created 2026-10-21T00:14:48Z; 148 likes / 548 downloads in the 2026-10-24 API check) tags license apache-2.0. The LICENSE file is Apache License 2.0. The card shows an extra_gated request form; the API gated field was false the same day. README: request access once (approval is automatic), then hf auth login. Sibling GGUF / NVFP4 / MLX / DFlash2 cards also tag apache-2.0. Hosted API and playground are not Apache-2.0. Not legal advice.
Is this a final model?
No. The README calls it a preview release. Specs and results may change before a final version. Marketing name is 30B-A4B; the architecture table prints 28.8B total / 4.3B activated.
Can I run it on a single GPU?
The card says NVFP4 enables inference on a single NVIDIA RTX 5090 (32 GB). GGUF, MLX, and DFlash2 draft siblings are published. Those are vendor size labels. forkalt has not loaded any Raon checkpoint.
Does it replace ChatGPT?
No sourced bake-off. The card positions Korean conversation and agentic tasks. See ChatGPT alternatives. Compact local coding MoE already on this site: Flint.
Sources & verification
- https://huggingface.co/KRAFTON/Raon-1-mini-preview
- https://huggingface.co/KRAFTON/Raon-1-mini-preview/blob/main/LICENSE
- https://huggingface.co/KRAFTON/Raon-1-mini-preview-gguf
- https://huggingface.co/KRAFTON/Raon-1-mini-preview-nvfp4
- https://huggingface.co/KRAFTON/Raon-1-mini-preview-mlx
- https://huggingface.co/KRAFTON/Raon-1-mini-preview-dflash2
- https://github.com/krafton-ai/Raon-1-mini-preview
- https://raon-mini.krafton.ai/en/terms
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.