KRAFTON Raon-1 mini explained: Korean, agents and single-GPU deployment in a 30B-class open-weight model

The takeaway

KRAFTON AI has released Raon-1 mini as a preview, not a final model. This 30B-class MoE language model emphasizes Korean conversation, instruction following, agents and inference efficiency rather than the largest possible parameter count.

Three things merit attention: selected reasoning and coding results against similar models, an official speculative-decoding option, and quantized formats for local deployment. There are also clear weaknesses. Neither “leads every benchmark” nor “runs smoothly on every computer” is supported.

This analysis uses the official model card and experience site. TboxIO has not independently rerun benchmarks, measured throughput or validated single-GPU deployment. Information checked October 25, 2026. The publisher warns that specifications and results may change before the final release.

What do 30B and A4B mean?

Specification	Official description
Total parameters	28.8B
Active parameters per token	4.3B
Architecture	Hybrid-attention MoE
Layers	25: 18 linear-attention and 7 full-attention layers
Context window	262,144 tokens
Active experts per token	6 routed plus 1 shared
Release	Preview, not final
License	Apache 2.0

Activating a subset reduces computation. It does not mean only 4.3B parameters need to be stored. Memory planning must include all weights, quantization, caches and runtime overhead.

Most layers use fixed-size states, while a smaller group retains KV caches. This explains the efficiency direction, not a promise of constant total memory or latency for arbitrarily long input.

Official scores: strengths and weaknesses

These are vendor-reported comparisons, not TboxIO measurements. Values follow the original table; different benchmarks should not be averaged together.

Benchmark	Raon-1 mini preview	Qwen3.6-35B-A3B	Gemma 4 26B-A4B
MMLU Pro	76.5	85.9	82.9
GPQA Diamond	84.3	83.2	80.7
AIME 2026	93.3	92.9	87.1
IFBench	84.2	66.6	80.9
SWE-Verified	70.8	63.2	61.6
Terminal-Bench Lite	47.0	63.3	52.0
BrowseComp Plus	55.7	34.8	32.2
BFCL v4 Overall	59.8	64.7	65.8

The results suggest a candidate for selected reasoning, instruction-following, repository repair and search tasks. However, Raon trails both comparison models on MMLU Pro, Terminal-Bench Lite and BFCL v4.

The contrast between SWE-Verified and Terminal-Bench Lite matters: better repository-fix results do not establish stronger terminal execution generally. Choose by workload rather than the most flattering row.

The full official table also includes dense models. Qwen3.8-27B, for example, scores 89.7 on GPQA Diamond versus Raon’s 84.3. Leading some listed MoE comparisons is different from outperforming every model. We do not invent a previous Raon version’s scores or an undisclosed generational improvement.

Evaluation conditions matter

SWE-Verified uses mini-SWE-agent. The card names gpt-5.6-luna as judge and user simulator. AutomationBench uses the public split, not the official evaluation set; GDPVal is a judge score, not Elo.

These conditions limit cross-source comparisons. Do not subtract Terminal-Bench Lite from another Terminal-Bench version reported in our earlier articles, or combine different harnesses into a single ranking.

Korean is a deliberate focus

The card lists Korean, English, Japanese, Chinese, Spanish, French and Thai among its training languages. Its unpublished internal Chatbot Korean test reports Raon 92.9, Qwen3.6-35B-A3B 74.5 and Gemma 4 26B-A4B 90.6.

That is a useful selection signal, not “92.9% everyday Korean accuracy.” Test honorifics, informal speech, consistent personas and mixed Korean-English inputs yourself. Inclusion of Chinese in training does not prove equal performance in Chinese business tasks.

What does up to 2.18× faster mean?

The accompanying DFlash2 draft model proposes candidate tokens for verification by the target model, reducing some sequential generation work.

The reported 2.18× example uses one NVIDIA B300, GSM8K and concurrency 1: autoregressive decoding produces 296.3 output tokens/second versus 646.3 with DFlash2. The test uses an MXFP8 target and specified settings including vLLM 0.28.0.

This is a vendor result for particular hardware and workloads, not a guarantee on other GPUs or a universal speed ranking against competitors. Aggregate throughput under higher concurrency is also different from each user’s output speed. Measure time to first token, individual generation speed, concurrent throughput and complete agent-task latency separately.

Can it run on an RTX 5090 or a Mac?

The publisher provides MXFP8 and NVFP4 options and says NVFP4 enables inference on a single NVIDIA RTX 5090 with 32GB VRAM. The card lists Transformers, vLLM, llama.cpp and MLX entry points.

Those are deployment paths, not evidence that every format, full context and high concurrency fit the same machine. For a Mac, check usable memory, the actual quantized file, runtime support and context length. Active parameter count alone is not a memory requirement.

Start with the platform-appropriate format, shorter context and concurrency 1, then increase load while recording memory and latency.

Agent integration

The card lists Hermes Agent, Pi and OpenCode via an OpenAI-compatible endpoint. API compatibility alone does not establish correct tool-call or reasoning-field handling: test message formats, tool-result feedback and stopping conditions.

The model is designed for thinking mode; disabling it is not recommended. Recommended sampling differs between coding/math and other tasks. Begin with the official configuration before changing settings one at a time.

Candidate trials include Korean support drafts, verifiable information organization, small repository fixes and constrained tool tasks. Track accepted completions, manual edits and total resources rather than response length.

Access and cost

The official site offers a guest experience and an API path involving application, invitation and registration. It says API access is free for a limited time, with possible per-user token and request limits. Future charges and policies will be announced separately.

This is not a permanent-free promise or a published per-million-token commercial tariff. Self-hosting still incurs hardware, operation and maintenance costs.

The official Hugging Face repository declares Apache 2.0, while the card notes gated access requests. Check access, format and license terms before downloading; open weights do not remove usage conditions.

Our judgment

Raon-1 mini preview is worth testing for teams emphasizing Korean interaction and control over deployment, with capacity to validate their own tooling. Its appeal combines deployable scale, selected task results and a paired acceleration option.

For broad knowledge, difficult terminal work or stable production service, include the reported weaknesses and preview status in the decision. Compare identical tasks, tool permissions and budgets before expanding traffic.

Sources

• Official KRAFTON model card: architecture, comparisons, evaluation setup, decoding and deployment.
• Official Raon site and FAQ: access and limited-time free service conditions.

Preview specifications, results and policies may change. Check current official information when integrating.