28.8B total / 4.3B active
262,144 tokens
Apache-2.0 for downloadable weights, configuration files, and model components; hosted API and playground use separate terms (official ModelScope mirrors currently expose generic "other" license metadata)
Preview release of a Korean- and English-capable reasoning and agentic language model designed and trained from scratch by KRAFTON AI. Raon-1 mini is a sparse Mixture-of-Experts model with 28.8B total parameters and 4.3B activated per token, described by KRAFTON as a 30B-A4B-class model. Its 25-layer hybrid architecture contains 18 fixed-state Raon Delta Attention layers and seven full-attention Gated MLA layers in a repeating 3:1 pattern. Raon Delta Attention modifies Kimi Delta Attention by decoupling the erase and write gates, while Raon-mHC reduces the computational cost of DeepSeek-style manifold-constrained hyper-connections and maintains four residual streams. Only the seven MLA layers retain a KV cache, reported at approximately 8 KiB per token in BF16. The model uses 128 routed experts, activates six routed experts plus one shared expert per token, has a 2,560-dimensional hidden state, 40 attention heads, an untied 259,840-token vocabulary, SiTU-GLU activations, and no positional encoding. It natively supports 262,144 combined input and output tokens, with documented output limits up to 32,768 tokens. The release is designed for thinking mode and returns reasoning separately from its final answer. Its chat template supports system and user messages, multi-step tool calls, tool responses, optional retention of earlier reasoning, and both thinking and non-thinking generation. KRAFTON reports scores of 84.3 on GPQA Diamond, 93.3 on AIME 2026, 84.2 on IFBench, 81.6 on LiveCodeBench v6, 70.8 on SWE-Verified using mini-SWE-agent, 55.7 on BrowseComp Plus, 59.8 on BFCL v4 Overall, and 67.3 on AA-LCR. Its internally developed Korean conversational benchmark score is 92.9; that benchmark is not publicly released. The publisher used gpt-5.6-luna as judge and user simulator for the reported evaluations, and its AutomationBench result uses the public rather than official evaluation split. The main approximately 32.70 GB checkpoint stores routed experts in MXFP8. An approximately 17.54 GB NVFP4 release is documented for single-GPU inference on a 32 GB RTX 5090, and a similarly sized MLX package supports Apple silicon. Official llama.cpp GGUF variants include IQ4_XS at approximately 16.07 GB, Q4_K_M at 19.69 GB, Q5_K_M at 22.18 GB, Q6_K at 25.84 GB, and Q8_0 at 30.69 GB; Q6_K and smaller use an importance matrix. A separate 2.71 GB DFlash2 draft model provides speculative decoding, with KRAFTON reporting up to 2.18× higher output throughput than autoregressive decoding on one NVIDIA B300. Documented inference paths include Transformers, KRAFTON's vLLM and llama.cpp branches, MLX, Ollama, Hermes Agent, Pi, and OpenCode. All preview repositories use automatic gated access, with one accepted request covering the family. This is explicitly a preview rather than a final model, and KRAFTON states that specifications and results may change.