3.1B main / 1.5B effective (dense; plus a 55M MTP module)
262,144 tokens; the LiteRT-LM bundle is limited to 4,096 tokens
Apache 2.0
Open-weight Korean- and English-capable instruction model designed for natural conversation, long-context instruction following, reasoning, and tool calling on memory-constrained and on-device hardware. Raon-1 nano is a dense model with 3.1 billion logical parameters but approximately 1.5 billion effective parameters; roughly 1.5B of its parameters reside in per-layer token-embedding tables. It was derived from the sparse Raon-1 mini preview through KRAFTON's MoE-to-Dense pruning and knowledge-distillation process, then continued-pretrained on three trillion tokens and post-trained through supervised fine-tuning, on-policy distillation from Raon-1 mini preview, and reinforcement learning. The 24-layer, 1,536-dimensional backbone contains 18 fixed-state Raon Delta Attention layers and six full-attention Gated MLA layers in a repeating 3:1 pattern. Raon Delta Attention uses separate erase and write gates, while the full-attention layers use 512-rank query and key-value projections. Only the six MLA layers retain a KV cache, requiring approximately 7 KiB per token in BF16. Other architectural features include four-stream Raon-mHC residual connections, 6,144-dimensional SwiGLU feed-forward layers, a 259,840-token vocabulary, tied input and output embeddings, no positional encoding, and a distinct 256-dimensional token-embedding table at every transformer layer. The native context window is 262,144 tokens. Thinking is enabled by default and KRAFTON recommends leaving it enabled; the chat template also supports direct-answer operation, retention or removal of historical reasoning, multi-step tool calls, and tool responses. Publisher-reported scores include 81.1 on MMLU-Redux, 50.9 on KMMLU-Redux, 44.2 on HAE-RAE 1.1, 65.2 on CLIcK, 75.9 on IFBench, 40.5 on MultiChallenge, 44.4 on FACTS Grounding, 38.8 on PersonaMem v2, 44.0 on PrefEval, 56.3 on BFCL v4, 36.0 on AndroidControl, 58.0 on NL2SH-ALFA, and 34.5 on BrowseComp Plus. KRAFTON additionally reports 49.2 on its unreleased Korean Office Work Bench and 91.0 on its unreleased Korean conversational benchmark. A separate 54.8M-parameter Multi-Token Prediction layer was trained during supervised fine-tuning and released for experimental speculative decoding; the target model subsequently underwent reinforcement learning without updating that MTP module. Native MTP support is available experimentally in KRAFTON's vLLM fork and through merged `_MTP` GGUF files. Official GGUF packages include Q4_K_M at approximately 2.08 GB, Q6_K at 2.65 GB, Q8_0 at 3.39 GB, and F16 at 6.27 GB, with slightly larger MTP-enabled counterparts. The four-bit MLX checkpoint occupies approximately 1.81 GB and includes fused Metal kernels plus a separate 62 MB MTP module. The approximately 1.82 GB LiteRT-LM bundle uses four-bit weights with FP16 activations, runs without thinking, and is limited to a 4,096-token context. Documented inference paths include Transformers, KRAFTON's vLLM and llama.cpp branches, MLX, LiteRT-LM, CUDA, Metal, Vulkan, and CPU. All released formats use Apache 2.0.