TL;DR
- KRAFTON AI released a preview of Raon-1 mini, a 30B-A4B Mixture-of-Experts model trained from scratch and published under Apache 2.0.
- Only 4.3B of its 28.8B parameters activate per token, and the context window stretches to 262,144 tokens.
- NVFP4 weights target a single 32 GB RTX 5090, and a DFlash2 draft model reportedly lifts output throughput by up to 2.18× on one NVIDIA B300.
- That 2.18× figure is a vendor benchmark, not an independent result.
Raon-1 Mini Preview: Small Active Footprint, Big Context
KRAFTON AI has put out a preview of Raon-1 mini, and the Hugging Face repository is blunt about what it is. In its own words, “Raon-1 mini is a 30B-A4B Mixture-of-Experts (MoE) language model, designed and trained from scratch by KRAFTON AI.” The target is Korean conversation and agentic tasks.
The numbers are specific. The model carries 28.8B total parameters, but only 4.3B are active for any given token. Each token routes through 6 experts, plus 1 shared expert, drawn from a pool of 128. Context length tops out at 262,144 tokens, and the license is Apache 2.0.
Two extras ship with it. The first is quantized weights in NVFP4 format, meant to support inference on consumer GPUs such as a single NVIDIA RTX 5090 with 32 GB of VRAM. The second is a DFlash2 draft model for speculative decoding. The repository says: “In our benchmarks on a single NVIDIA B300, it delivers up to 2.18× higher output throughput than autoregressive decoding.” MXFP8 weights are part of the release too.
The Draft Model Is the Real Story
Plenty of labs ship a 30B-class MoE. Fewer ship the speed machinery in the same drop. That’s what caught my eye here, because a model nobody can serve cheaply is a model nobody tests.
Think of speculative decoding as a lumber yard. An apprentice saws rough planks quickly, and the master carpenter runs a hand along the stack, pulling out the warped ones instead of cutting every plank himself. The draft model is the apprentice. The main model is the master. When the apprentice guesses well, the yard moves faster.
Does that explain 2.18×? Not by itself. The figure is “up to,” it comes from KRAFTON’s own tests on a specific workload, and it was measured on a B300, a datacenter part most readers won’t have in a closet. Treat it as a ceiling the vendor reached, not a promise about your setup.
The RTX 5090 angle matters for a different reason. A 32 GB consumer card is something a hobbyist, a small startup or a university lab can actually buy. If NVFP4 weights run acceptably there, a Korean-focused agentic model stops being something you rent and becomes something you can poke at on your desk. I’d call that the more interesting claim, and it’s also the one with the least published evidence so far.
Then there’s the competition. KRAFTON benchmarks the preview against Qwen3.6-35B-A3B, Gemma 4 26B-A4B, Qwen3.8-27B and Muse Glimmer 30B across reasoning, instruction following and agentic tasks. By name, Gemma 4 26B-A4B sits in the same active-parameter class as Raon-1 mini. That’s a pointed choice of opponent. But the material I have doesn’t include the score tables, so I can’t tell you who won, and I won’t guess.
A Korean-First Model Built on Bilingual Ambitions
Raon-1 mini isn’t a one-off. KRAFTON AI has been developing the Raon family of open models with bilingual Korean and English capability in mind, along with agentic workflows and efficient inference. The architecture reflects that last goal, using hybrid linear and full attention mechanisms.
That combination makes sense for the context length on offer. Full attention over 262,144 tokens is expensive, and hybrid designs exist to blunt that cost. Whether this particular mix holds quality at long range is a question the repository’s numbers, not my reading of an architecture note, will have to answer.
The word “preview” deserves attention too. It signals something unfinished. Weights, behavior and benchmark results may change before whatever comes next, so anyone building on it should pin versions and expect movement.
And the open license counts. Apache 2.0 is permissive, which means companies can fine-tune and ship products on it without a negotiation. For a language where open, high-capability options are thinner on the ground, that’s a practical gift, whatever the final benchmark tables say.
Three Checks Before You Trust the Preview
First, look for independent throughput numbers. The 2.18× claim needs someone outside KRAFTON to reproduce it, ideally on more than one workload and on hardware other than a B300. If third-party results land well below it, that tells you how workload-specific the gain is.
Second, watch for real-world reports from RTX 5090 owners running the NVFP4 weights. Single-card serving is the headline accessibility claim, so speed, memory headroom and output quality at long context are what to look for.
Third, read the benchmark comparisons against Qwen, Gemma and Muse Glimmer closely once people dig into them. Self-selected tasks flatter the home team, so the useful signal will be Korean conversation and agentic tests run by someone else.
Editor's Note
I care less about the 2.18× headline than about the RTX 5090 claim. Speedups on a B300 are a datacenter story, and most people reading this will never touch one. A 30B-class model that a hobbyist can actually run is a different kind of release. I'm waiting for someone outside KRAFTON to post real Korean-language results from a single consumer card. That's the test I'd trust.
– Sanket Chaukiyal, founder, SmartChunks
FAQ
What is Raon-1 mini?
It's a preview of a 30B-A4B Mixture-of-Experts language model from KRAFTON AI, trained from scratch for Korean conversation and agentic tasks. It has 28.8B total parameters with 4.3B active per token, and it's released under Apache 2.0.
Can it run on a consumer GPU?
The NVFP4 quantized weights are meant to support inference on consumer cards such as a single 32 GB NVIDIA RTX 5090. The material reviewed doesn't include independent speed or quality results for that setup.
What does the DFlash2 draft model do?
It supports speculative decoding. KRAFTON reports up to 2.18× higher output throughput than autoregressive decoding on a single NVIDIA B300. That's a self-reported benchmark on specific workloads and hardware.
How does Raon-1 mini compare with rival models?
KRAFTON benchmarks it against Qwen3.6-35B-A3B, Gemma 4 26B-A4B, Qwen3.8-27B and Muse Glimmer 30B on reasoning, instruction following and agentic tasks. The specific scores weren't part of the verified material, so no winner can be named.
