🌏 中文版
Model Info
| Field | Value |
|---|---|
| Model ID | KRAFTON/Raon-1-mini-preview |
| Vendor | KRAFTON AI (AI research arm of KRAFTON, the studio behind PUBG) |
| Parameters | 28.8B total / 4.3B active (30B-A4B class MoE; 6 of 128 routed experts plus 1 shared expert active per token) |
| Context Window | 262,144 tokens |
| Input Pricing (USD/1M tokens) | Not available (open-weight preview release; no official or third-party hosted API pricing yet) |
| Output Pricing (USD/1M tokens) | Not available (same as above; the playground is free for a limited time but that isn't API pricing) |
| Open Source | Yes (Apache 2.0 for the weights; the hosted API/playground is governed by a separate service agreement outside Apache 2.0) |
| Release Date | 2026-10-21 (preview) |
| Official Announcement | KRAFTON/Raon-1-mini-preview (Hugging Face model card) |
| HuggingFace | KRAFTON/Raon-1-mini-preview |
| Family | Raon (KRAFTON AI's open-model brand, launched April 2026 with a speech LLM, a real-time voice model, a TTS model, and a vision encoder; Raon-1 mini is the family's first general-purpose text LLM) |
Highlights
- Korean conversation: scores 92.9 on the internal Chatbot Korean benchmark, well ahead of the similarly-sized MoE model Qwen3.6-35B-A3B (74.5) and also beating Gemma 4 26B-A4B (90.6)
- Hybrid attention architecture: of 25 layers, 18 use KRAFTON's own Raon Delta Attention (linear attention with a fixed-size state) and 7 use Gated MLA (which keeps a KV cache), bringing per-token KV cache down to about 8 KiB (BF16) while natively supporting 262,144 tokens of context
- Built-in speculative decoding: ships with a DFlash2 draft model that delivered up to 2.18x higher output throughput on a single NVIDIA B300 across several benchmarks
- Lower edge-deployment bar: an NVFP4-quantized build runs entirely on a single NVIDIA RTX 5090 (32GB VRAM); MXFP8, GGUF and MLX builds are also available
Benchmark Results
| Benchmark | Score | Previous Gen | Best MoE Peer |
|---|---|---|---|
| Chatbot Korean (internal) | 92.9 | — (first model) | Gemma 4 26B-A4B 90.6 (Qwen3.6-35B-A3B only 74.5) |
| SWE-bench Verified | 70.8 | — (first model) | Qwen3.6-35B-A3B 63.2 (for reference: dense Qwen3.8-27B 74.6) |
| AIME 2026 | 93.3 | — (first model) | Qwen3.6-35B-A3B 92.9 |
| MMLU Pro | 76.5 | — (first model) | Qwen3.6-35B-A3B 85.9 (Raon-1 mini noticeably trails on general knowledge) |
| BFCL v4 Overall (tool calling) | 59.8 | — (first model) | Qwen3.6-35B-A3B 64.7 (for reference: dense Qwen3.8-27B 72.4) |
⚠️ All figures are KRAFTON's own measurements (judged by gpt-5.6-luna) and await independent reproduction. Chatbot Korean is an unreleased internal benchmark that can't be verified directly by third parties. KRAFTON also notes this is a preview — specs and scores may change before a final release.
Source — KRAFTON's own measurements on its official Hugging Face model card (judge model gpt-5.6-luna); the Korean conversation score is from an unreleased internal benchmark. Y axis runs 0–100, baseline y=300, plot top y=60, 2.4 units per point (y = 300 − score × 2.4).
vs. Previous Gen / Competitors
Raon-1 mini is the Raon family's first general-purpose LLM, so there's no "previous generation" to compare against — the relevant comparison is other MoE models of similar size. Against Qwen3.6-35B-A3B, Raon-1 mini wins clearly on Korean conversation (92.9 vs. 74.5) and SWE-bench Verified (70.8 vs. 63.2), but trails on MMLU-Pro (76.5 vs. 85.9) and BFCL tool calling (59.8 vs. 64.7). That split suggests KRAFTON optimized training for Korean dialogue quality and software-engineering agent tasks rather than broad world-knowledge coverage.
On pricing, KRAFTON has gone fully open-weight plus a time-limited free playground, with no hosted API pricing announced — the reverse of how most closed-source labs operate (price first, open-source older versions later). It reads more like a one-time technical signal meant to win developer attention than an immediate monetization push.
What's arguably more notable is who released it: KRAFTON is a game studio, not a traditional AI lab, yet it chose to train a 30B-class MoE foundation model from scratch rather than simply calling or fine-tuning an existing open model. That suggests the bar for training your own foundation model has dropped low enough that a large company in a vertical industry like gaming now finds it worthwhile.
What This Means for Agent Development
- If you're building a customer-support or in-game NPC dialogue agent that needs natural Korean: Raon-1 mini's lead on Korean benchmarks makes it one of the few open models explicitly tuned for Korean conversational tone, sparing you the work of forcing a larger model into the right register.
- If you're building an edge-deployed or self-hosted, low-latency agent: the NVFP4 build runs on a single RTX 5090, and DFlash2 speculative decoding adds up to 2.18x throughput — a good fit if you want to avoid cloud API dependency while keeping response times tight.
- Not a good fit: tasks that need broad world knowledge or complex multi-tool chains, where Raon-1 mini is the weakest in its class on both MMLU-Pro and BFCL. Use Qwen3.6-35B-A3B or a larger dense model there instead. And since this is still a preview, don't bet production workloads on specs or scores that may still change before the final release.
Today's Takeaway
I used to assume the race to train foundation models from scratch was confined to traditional AI labs and Chinese vendors. A game studio training and open-sourcing a 30B-class MoE model — designing its own attention variant (Raon Delta Attention) and training it from the ground up, not just wrapping someone else's API — shows that the bar for "who gets to build foundation models" has dropped. Large companies in vertical industries can now step into this layer too, and apparently want to.
References
- KRAFTON/Raon-1-mini-preview (Hugging Face model card)
- KRAFTON AI official site
- Raon-1 mini playground
- Seoul Economic Daily: Krafton Releases Raon-1 Mini, Its In-House AI Language Model
- Digital Today: Krafton unveils 30B-class LLM 'Raon-1 Mini' to boost cost efficiency
- Seoul Economic Daily: Krafton Launches AI Model Brand 'Raon,' Releases Four Open-Source Models