Is Raon-1 mini Free? Playground Access and Licensing Explained
KRAFTON's Raon-1 mini offers a free limited-time playground and Apache 2.0 open weights. Here's how access and licensing actually work.
Is Raon-1 mini free to use?
Yes, in two separate ways. KRAFTON AI runs a free, time-limited web playground where anyone can chat with Raon-1 mini preview without installing anything. Separately, the model weights are published on Hugging Face under the Apache 2.0 license, which means developers can download, run, modify, and build commercial products on top of it with no royalty or licensing fee. The playground access is temporary; the open weights are not.
TL;DR
- The playground is a limited-time offer, hosted by KRAFTON AI, letting anyone try Raon-1 mini preview in a browser with no setup or account requirements mentioned in the model card.
- The underlying model is released under Apache 2.0, a permissive open-source license that allows commercial use, modification, and redistribution without paying KRAFTON.
- Raon-1 mini is a 30B-A4B Mixture-of-Experts model with 28.8B total parameters and 4.3B activated per token, trained from scratch by KRAFTON AI.
- It’s explicitly a preview release, so KRAFTON has flagged that specs and benchmark results may change before a final version ships.
- It supports quantized inference down to a single RTX 5090, with NVFP4 quantization cited as enabling local use on a 32GB consumer GPU.
- The model targets Korean conversation and agentic tool-use, while also handling English, Japanese, Chinese, Spanish, French, and Thai.
- Benchmark tables on the model card show it competitive with or ahead of similarly sized MoE models like Qwen3.6-35B-A3B and Gemma 4 26B-A4B on several reasoning and knowledge tests.
Remy doesn't build the plumbing. It inherits it.
Other agents wire up auth, databases, models, and integrations from scratch every time you ask them to build something.
Remy ships with all of it from MindStudio — so every cycle goes into the app you actually want.
What exactly is Raon-1 mini?
Raon-1 mini is an open-weight language model built by KRAFTON AI, the games company behind PUBG, as part of its Raon model family. It’s a Mixture-of-Experts (MoE) design described as “30B-A4B,” meaning it has roughly 30 billion total parameters but only activates about 4 billion per token. The precise figures on the model card are 28.8B total parameters and 4.3B activated parameters, with 128 routed experts (6 active per token) plus one shared expert.
The architecture uses a hybrid attention layout: 18 of 25 layers run Raon Delta Attention (RDA), a linear-attention variant KRAFTON says improves on Kimi Delta Attention for long-context handling, while the remaining 7 layers use full (gated MLA) attention with a KV cache. KRAFTON also built a custom residual-connection method called Raon-mHC, meant to cut the computational overhead of DeepSeek’s manifold-constrained hyper-connections and speed up training.
The model natively supports context windows up to 262,144 tokens and comes with a DFlash2 draft model for speculative decoding, plus MXFP8 and NVFP4 quantized formats. According to the model card, NVFP4 quantization makes it possible to run inference on a single NVIDIA RTX 5090 with 32GB of VRAM, putting it within reach of serious hobbyist and small-team hardware rather than requiring a data center GPU.
KRAFTON designed Raon-1 mini around Korean conversational ability and agentic execution, including tool calling compatible with frameworks like Hermes Agent, Pi, and OpenCode. It was trained across seven languages: Korean, English, Japanese, Chinese, Spanish, French, and Thai, with Korean and English cited as its strongest areas.
How does the free playground work?
The playground is a hosted web chat interface at KRAFTON’s own domain, separate from the Hugging Face repository. The model card describes it plainly: try Raon-1 mini “without setting anything up,” chatting with it “for free, for a limited time.” That phrasing matters for two reasons.
First, it’s a convenience layer. You don’t need a GPU, you don’t need to download multi-gigabyte weight files, and you don’t need to configure an inference server. It’s meant for people who want to evaluate the model’s Korean conversation quality or agentic behavior before committing engineering time to self-hosting it.
Second, “limited time” means exactly what it says. KRAFTON hasn’t published a specific end date in the model card, but the wording signals this is a promotional or evaluation window tied to the preview release, not a permanent free tier. Anyone planning to build a workflow or demo around the hosted playground should treat it as temporary and plan to migrate to self-hosted inference (or a future KRAFTON product offering) once the free window closes.
The playground also doesn’t change the licensing status of the model itself. Using the web chat doesn’t grant you any rights to the weights, and conversely, not using the playground doesn’t block you from downloading and running the model yourself under Apache 2.0.
What does the Apache 2.0 license actually permit?
- ✕a coding agent
- ✕no-code
- ✕vibe coding
- ✕a faster Cursor
The one that tells the coding agents what to build.
Apache 2.0 is one of the most permissive open-source licenses in common use for AI models, and KRAFTON applied it to the full Raon-1 mini release. In practical terms, that means:
- You can download the weights from Hugging Face and run them locally or on your own servers.
- You can fine-tune, modify, or merge the model into other systems.
- You can use it commercially, including in paid products, without owing KRAFTON licensing fees.
- You can redistribute modified versions, provided you retain the required license and attribution notices.
This is a meaningfully different posture from licenses that restrict commercial use or cap usage by company size or revenue (a pattern seen with some other open-weight model releases). Apache 2.0 carries no such restrictions. The tradeoff is that Apache 2.0 offers no warranty and no liability coverage from KRAFTON, standard for this license family, so production use is entirely at the deployer’s own risk.
Because this is a “preview” release, though, the license covering today’s weights may not automatically carry forward unchanged specs. KRAFTON states outright that “specs and results may change before the final version,” which is a caution about model behavior and benchmarks, not a license caveat, but it’s worth reading as a signal that the current checkpoint is a snapshot, not necessarily the long-term reference implementation.
Is Raon-1 mini worth running yourself?
For teams already working with Korean-language or multilingual applications, the case is straightforward: it’s a from-scratch MoE model with benchmark numbers that the model card shows as competitive with other open MoE models in its size class on tests like MMLU Pro, GPQA Diamond, and AIME-style math reasoning. The NVFP4 quantization path that fits on a single RTX 5090 lowers the hardware bar considerably compared to needing multi-GPU server racks.
The practical downsides are the ones inherent to any preview release: benchmark numbers and model behavior could shift before a stable version ships, documentation and community tooling support are still maturing (the Hugging Face listing shows modest download and like counts so far), and anyone deploying it in production should expect to re-evaluate when KRAFTON issues updates.
For quick evaluation without any infrastructure commitment, the free playground is the lower-friction option. For anything that needs to run indefinitely, integrate into a pipeline, or avoid dependency on KRAFTON’s hosting decisions, pulling the Apache 2.0 weights directly is the more durable path.
Frequently Asked Questions
Do I need an account to use the Raon-1 mini playground?
The model card doesn’t mention an account requirement, describing the experience simply as chatting with the model “for free” through the hosted link. Specific sign-up steps aren’t detailed in KRAFTON’s documentation.
Can I use Raon-1 mini commercially?
Yes. The model is released under Apache 2.0, which permits commercial use, modification, and redistribution without a royalty owed to KRAFTON, subject to standard license attribution requirements.
What hardware do I need to run Raon-1 mini myself?
KRAFTON provides quantized versions in MXFP8 and NVFP4 formats. The NVFP4 version is described as capable of running on a single NVIDIA RTX 5090 with 32GB of VRAM, which is notably lower than what many models of comparable total parameter count require.
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
Is this the final version of Raon-1 mini?
No. KRAFTON explicitly labels it a preview release and states that specifications and benchmark results may change before a final version is published.
What languages does Raon-1 mini support well?
It was trained on Korean, English, Japanese, Chinese, Spanish, French, and Thai, with the model card citing particularly strong Korean performance alongside solid English capability.
