A 30-billion-parameter model no longer has to be a datacenter-only workload. KRAFTON AI has published Raon-1 mini preview, a Mixture-of-Experts model with 30B total and 4B active parameters, and it points to a concrete deployment path: with NVFP4 quantization, inference can run on a single NVIDIA RTX 5090 with 32 GB of VRAM. That is not just a spec-sheet detail. It is where the open LLM race shifts from parameter counts to hardware practicality.

Raon-1 mini is trained from scratch, not a fine-tuned version of an existing base model. KRAFTON positions it around Korean-language conversation and agentic tasks, with stated support for tools such as Hermes Agent, Pi, and OpenCode. The Apache 2.0 license removes many distribution and modification constraints, which matters for teams that want to keep control of the execution environment.

The package also includes a DFlash2 draft model for speculative decoding, a technique that accelerates inference by letting a smaller model propose tokens and using the main model to verify them. Combined with a MoE that has a low number of active parameters, the practical effect is less pressure on VRAM and lower perceived latency, especially on local hardware.

The most telling choice is the publication of NVFP4 and MXFP8 quantization formats. NVFP4 is a 4-bit format tied to recent NVIDIA GPUs; the model card explicitly links it to the 32 GB RTX 5090. This does not mean Raon-1 mini is a consumer model in the strict sense: 30B total parameters still require meaningful memory, and aggressive quantization brings trade-offs in response quality and stability that need to be tested in practice.

For self-hosted deployments, the structural message is clear: open model providers are starting to ship not only weights but also a serving path. A draft model and quantization formats are delivered alongside the checkpoint, reducing the infrastructure adaptation work. That changes incentives: evaluating an agentic LLM on owned hardware becomes more realistic without relying on cloud APIs, especially for language or business data that organizations prefer to keep in place.

There is also a second, less visible effect. Adopting NVFP4 as the main access route ties deployment to a specific NVIDIA GPU line. That is a de facto constraint, even though the model license remains open. MXFP8 or other formats do not remove dependence on a hardware ecosystem; they shift it. For teams evaluating on-premise use, the cost of an RTX 5090 may be more predictable than cloud billing, but TCO includes maintenance, power, and replacement.

Raon-1 mini preview also signals maturity in the open model supply: a company known primarily for gaming is investing in an LLM trained from scratch, with a Korean and agentic focus, and releasing it under a permissive license. It is not a generalist model chasing global benchmarks; it is an object with a precise positioning, built to be run rather than just downloaded. In that sense, the story is not only the model itself, but the fact that the deployment path has become part of the product.