A games studio built Raon-1 mini, an open-weight MoE that targets a single RTX 5090

Open-weight models usually come from AI labs, cloud companies, or chipmakers. Raon-1 mini comes from a games studio. KRAFTON, the Korean company behind PUBG, released a preview of it on 21 October 2026, a model it says it designed and trained from scratch. The novelty of the source is part of the story, but the more interesting part for anyone running models at home is what it is built to run on.

What was released

Raon-1 mini is a mixture-of-experts model with 28.8 billion total parameters and 4.3 billion active per token, which KRAFTON labels “30B-A4B”. It routes each token through 6 of 128 routed experts plus one shared expert. It supports a native context window of 262,144 tokens, and KRAFTON says it was trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, with Korean its strongest language and English described as solid.

The weights are released under Apache 2.0, with one caveat worth stating plainly: access is gated. You accept the model card’s conditions before downloading, an approval the card describes as automatic, and the hosted “Raon Service” is explicitly not covered by Apache 2.0. So these are genuinely open weights under a permissive licence, with a click-through gate on the files and separate terms on the company’s own hosted offering.

Built for one high-end card

What sets Raon-1 mini apart from the run of 30-billion-class releases is where it is aimed. KRAFTON documents a 4-bit quantised build, in the NVFP4 format, that it says runs inference on a single NVIDIA RTX 5090 with 32GB of VRAM, with a 32,768-token serving profile on that card. No measured memory use or throughput figures are published for that setup, so this is best read as the configuration the vendor targets rather than a benchmarked result. An MXFP8 build on a 32GB card is suggested to need CPU offload, so the single-card story really rests on the 4-bit version. A separate NVFP4 repository is published, and a third party has reported an MLX build for Apple Silicon through tools like MLX-LM.

It also ships with a companion DFlash2 draft model for speculative decoding, a technique where a small fast model proposes several tokens at once and the main model verifies them, which can raise throughput without changing the output. KRAFTON claims up to 2.18 times higher output throughput than plain decoding on a single NVIDIA B300, and notes the speed-up varies by hardware and workload. Shipping the drafter alongside the model is a practical touch for people who care about tokens per second on their own machine.

What the numbers do and do not say

KRAFTON’s self-reported benchmarks, dated to this 21 October preview, are strong where the model is clearly tuned to be strong. On its highlighted set it posts 84.3 on GPQA Diamond, 93.3 on AIME 2026, 84.2 on IFBench, and 70.8 on SWE-Verified, figures that would be respectable for a model this size if they hold up under independent testing.

The honest qualifier is that the model’s own published comparison table shows it trailing on other axes. On the BFCL v4 tool-calling benchmark it scores 59.8, below Qwen3.8 27B at 72.4 and other listed mixture-of-experts models. On Terminal-Bench Lite it scores 47.0, below Qwen3.6 35B-A3B at 63.3. On MMLU Pro it sits at 76.5, under the comparison models including Gemma 4 26B. So the “leading results” framing applies to maths, science and instruction-following, not to agentic and tool-use work, where it currently lags. These are the publisher’s own figures, not independent evaluations, and a preview’s numbers can move before a stable release.

Where it fits

A 30-billion-class mixture-of-experts model with a few billion active parameters is now a crowded part of the field, alongside releases like Qwen3.8, GLM-5.3 and the larger Reflection Beam. Raon-1 mini’s distinguishing pitch is the deliberate single-5090 target and the bundled speculative-decoding drafter, which make it a reasonable thing to try if you have a 32GB consumer card, want a reasoning-leaning model, and work partly in Korean or the other supported languages.

The cautions are the ones that come with any preview. KRAFTON says specifications and results may change, a technical report is still to come, and the strongest claims are its own. We are holding a catalogue entry until a stable release, when the licence gate, the quantisation options and the hardware figures can be re-confirmed against the source. For now it is a promising and slightly unexpected arrival, strong on reasoning, unproven on agentic work, and genuinely runnable on a single high-end card.