How Much To Run AI
Open weights

Raon-1 Mini Preview: Local Reasoning, 5090 Setup and Cost

Raon-1 mini preview is KRAFTON AI’s Apache 2.0 text reasoning and tool-calling model, with Korean conversation among its stated focuses. The card calls it 30B-A4B; its detailed table lists 28.8B total and 4.3B activated parameters. A documented NVFP4 and DFlash2 recipe targets one RTX 5090 32GB at a 32,768-token context. Choose that exact configuration before budgeting hardware.

Content reviewed:

Raon-1 mini preview routes: hosted playground, RTX 5090 with NVFP4 and DFlash2 at 32K, or GGUF with the KRAFTON llama.cpp branch.
Route guide. The 32K RTX 5090 profile belongs to its exact NVFP4, DFlash2 and vLLM settings; GGUF requires a separate measurement.

Use this model

Online access and trials

Official model repository

Use cases

  • Evaluate Korean text conversation or a reasoning task with a checkable answer.
  • Test structured tool calling with an agent client that executes permitted tools.
  • Compare a quantized local session with the hosted playground trial.

Separate checkpoint access from the playground trial

KRAFTON supplies model repositories and a browser playground. The card describes a one-time Hub access request with automatic approval across the preview family; follow that step if the repository prompts for it. The Apache 2.0 model license is separate from the playground’s free-for-a-limited-time offer. The trial does not establish an ongoing hosted API tariff.

The model returns text and supports reasoning and tool calls. A tool-capable response still needs an agent client to run the requested action and send results back. Choose a small coding, Korean-language or reasoning task that you can evaluate before comparing cost per completed task.

Use the specific 32K RTX 5090 profile

The official 5090 example uses KRAFTON/Raon-1-mini-preview-nvfp4 together with the Raon DFlash2 draft model. It requires CUDA Toolkit 13 with nvcc on PATH and the card’s supported vLLM installation. Its launch command sets max-model-len to 32768, max-num-seqs to 8, FP8 KV cache and a 3GiB KV-cache allocation, alongside the Raon reasoning and tool parsers. Retain the complete recipe; copying only the model name loses those constraints.

The architecture table’s 262,144-token context and the GGUF examples’ 256K allocation are separate from this 32K 5090 profile. Neither the 4.3B activated count nor the single-card example establishes that a full-precision checkpoint or a 256K session fits 32GB. Speculative decoding also loads a draft model, so measure the complete target-and-draft process.

ConfigurationWhat the card establishesBudget boundary
NVFP4 + DFlash2 · RTX 5090 32GBDocumented 32,768-token vLLM launch profileExact cache, concurrency, runtime and both models
Model context configuration262,144 tokensNo universal 32GB memory claim
GGUF / llama.cpp routeKRAFTON branch and Q4_K_M exampleDifferent format and runtime; measure independently

Build the matching llama.cpp branch for GGUF

The model card directs GGUF users to KRAFTON’s raon-1-mini-preview branch while upstream support is pending. Build that branch rather than assuming an arbitrary installed llama.cpp release understands the architecture. The card’s terminal and server examples select KRAFTON/Raon-1-mini-preview-gguf:Q4_K_M; starting them downloads the checkpoint.

Follow the official full launch options for reasoning, chat templates and sampling. For a server, the documented local endpoint is http://127.0.0.1:8000/v1 with alias Raon-1-mini-preview. The GGUF path can use optional DFlash2 decoding, but its timing and memory are not interchangeable with the NVFP4 vLLM profile.

Official GGUF build path; model launch is a separate next step
git clone -b raon-1-mini-preview https://github.com/krafton-ai/llama.cpp
cd llama.cpp
# Omit -DGGML_CUDA=ON on Apple Silicon.
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release -j

Measure complete reasoning and agent sessions

Reasoning output, prompt processing, tool execution, retries and startup all occupy the rented machine. For an illustrative single-GPU allocation of two hours at $0.80/GPU-hour, rental is $1.60 before storage or other charges. This example uses assumed inputs, not a Raon speed estimate or a provider quote.

For owned hardware, use whole-system power over the full session and add any allocated hardware cost. Compare accepted results under the same context and tool workload. The card’s B300 speculative-decoding speedup cannot be carried over to a 5090 or a different concurrency. This site has no measured Raon hardware profile or ongoing hosted token rate.

Before you use it

  • The 32GB 5090 recipe is specific to its NVFP4, draft-model and 32K settings; other formats and contexts need a separate memory measurement.

How to get started

Choose the playground, the documented 5090 profile or the matching GGUF runtime, then test one reproducible task and record the full session before scaling it.

Sources

Sources (3)

Calculators