MoE workstation
Raon-1 mini is a model that thinks before it answers and answers in Korean as easily as English. Picture a careful analyst who drafts an outline first and speaks second, at 4B-of-weights speed per token. It exists because game company KRAFTON needed believable AI teammates and shipped the capability as open weights instead of keeping it inside the studio. Reach for it when the work is Korean-first assistants, agentic search, or one-RTX-5090 deployments. The DFlash2 draft model ships beside it, so speculative decoding works on day one. Honest status: preview weights (NVFP4, MLX, GGUF) under Apache-2.0, gated but auto-approved; benchmarks vendor-measured (AA-Omniscience 55.7, GPQA Diamond 84.3, Chatbot Korean 92.9).
korean agentic thinking local moe apple-silicon nvidia
-
Parameters
-
30.0B -
Context
-
256k
-
License
-
apache 2.0
-
Origin
-
🇰🇷 South Korea -
Released
-
Oct 2026
Run it locally #
Per-quant memory needs and a static "can you run it?" reference - no rig entry required
NVFP4 16.0GB 17.0GB min 24.0GB rec
NVIDIA FP4 - Blackwell or Apple MLX, near-lossless
Q4_K_M 17.5GB 18.0GB min 24.0GB rec
Balanced - the usual local sweet spot
The reference hardware
22 reference configs, drawn in-house. Scroll for more.
Can you run it? - reference rigs
| Rig | NVFP4 | Q4_K_M |
| NVIDIA Jetson Orin NX 16GB | [no -> cloud](#cloud-pricing) | [no -> cloud](#cloud-pricing) |
| Single GTX 1080 Ti (11GB) | [no -> cloud](#cloud-pricing) | offload |
| 4x H100 80GB (320GB) | [no -> cloud](#cloud-pricing) | fast 3158.6t/s |
| NVIDIA DGX Station 748GB | [no -> cloud](#cloud-pricing) | fast 1885.7t/s |
| 4x RTX 5090 (128GB) | fast 1848.0t/s | fast 1689.6t/s |
| 8x RTX 3090 rack (192GB) | [no -> cloud](#cloud-pricing) | fast 1765.4t/s |
| AMD Instinct MI300X (192GB) | [no -> cloud](#cloud-pricing) | fast 1255.1t/s |
| 4x RTX 4090 (96GB) | [no -> cloud](#cloud-pricing) | fast 950.4t/s |
| 2x RTX 5090 (64GB) | fast 924.0t/s | fast 844.8t/s | | Single RTX 5090 (32GB) | fast 462.0t/s | fast 422.4t/s | | RTX PRO 6000 Blackwell (96GB) | fast 462.0t/s | fast 422.4t/s | | 2x RTX 3090 (48GB) | no -> cloud | fast 441.4t/s | | Mac Studio M4 Ultra 192GB | fast 307.1t/s | fast 280.8t/s | | Mac Studio M4 Ultra 512GB | fast 307.1t/s | fast 280.8t/s | | Single RTX 4090 (24GB) | no -> cloud | fast 237.6t/s | | MacBook Pro M5 Max 128GB | fast 172.7t/s | fast 157.9t/s | | Dual EPYC 9004 + 768GB DDR5-4800 | no -> cloud | fast 108.6t/s | | DGX Spark 128GB unified | fast 70.4t/s | fast 64.4t/s |
| Ryzen AI Max+ 395 128GB | [no -> cloud](#cloud-pricing) | fast 60.3t/s |
| Jetson AGX Orin 64GB | [no -> cloud](#cloud-pricing) | fast 48.3t/s |
| Epyc + 512GB DDR4-3200 + 2x RTX 3090 | [no -> cloud](#cloud-pricing) | fast 48.3t/s |
| Epyc + 512GB DDR4-2400 + 2x RTX 3090 | [no -> cloud](#cloud-pricing) | fast 36.2t/s |
Fit tiers use the same will-it-run logic as the rig finder. For comfortable fits, the badge reflects decode speed: fast >=20 t/s, ok 8-20 t/s, slow <8 t/s. t/s is a bandwidth estimate, not a measured benchmark.
PRICE HISTORY
Inference cost over time #
Data accumulates from the first daily sync - longer ranges populate over time. Prices come from OpenRouter snapshots, not a historical API.
price history...