KRAFTON, the South Korean game company behind PUBG, has quietly put its first open-weight language model on Hugging Face: Raon-1 mini preview. This is not a fine-tune of someone else's checkpoint — the model card describes it as "designed and trained from scratch by KRAFTON AI", a 30B-A4B sparse MoE under the Apache 2.0 license. A major game studio entering the base-model arena is, by itself, more interesting than any benchmark line.
A 30B stuffed with 2026's favorite components
The spec sheet: 28.8B total parameters, 4.3B activated per token — only 6 of 128 routed experts fire per token, plus 1 shared expert; 25 layers, hidden size 2,560. The attention design is the most interesting part: a 3:1 hybrid of 18 linear-attention layers (RDA) and 7 Gated MLA layers. Only those 7 MLA layers keep a KV cache, at roughly 8 KiB per token in BF16 — which is where the native 262,144-token context comes from.
Both custom-named components stand on other people's shoulders. Raon Delta Attention, per the official description, decouples the erase and write gates of Kimi Delta Attention to improve long-context performance. Raon-mHC is a cost-reduced take on DeepSeek's manifold-constrained hyper-connections, which the team says speeds up training. Add NoPE (no positional encoding) and a 4-stream residual, and this 30B remixes nearly every fashionable architecture trick of 2026 — the standard playbook for a small team: don't invent the wheel, just modify it beyond recognition.
The deployment story: fits on one RTX 5090
Quantization options are generous: official MXFP8 and NVFP4 builds, with the NVFP4 version running inference on a single NVIDIA RTX 5090 (32 GB VRAM). Aggregator theopenweights, reading the weights metadata, puts the FP8 build at about 31.9 GB and estimates the 4-bit build fits on 24 GB cards (RTX 3090/4090). A DFlash2 draft model ships alongside for accelerated speculative decoding. For latency-sensitive on-device and in-game applications, a 4.3B-active inference profile is a fairly pragmatic choice.
The officially claimed parts: apply a discount
The highlights on the model card are all vendor claims: built-in thinking with "leading results" on the math and science benchmarks shown; leads on instruction-following and search benchmarks; trained on Korean, English, Japanese, Chinese, Spanish, French and Thai, with the strongest Korean performance; tool calling for agentic workflows with support for Hermes Agent, Pi and OpenCode. Two caveats: the README states plainly that this is a preview and "specs and results may change", and the benchmark tables are self-reported — treat everything as provisional until the final release and independent reproduction. Current traction is modest: 249 downloads on Hugging Face, 3 stars on GitHub.
So what
theopenweights called it right: this is "more of a signal of intent than a drop-in replacement". The interesting question comes next — a company rooted in interactive entertainment has trained a MoE from scratch; which direction will it optimize toward? The published record gives no answer yet; the final release will. When the company behind PUBG starts training its own MoE, the guest list at the open-weights table turns out to be longer than most people assumed.