Kanana-2

30B total / 3B active MoE with MLA (Multi-head Latent Attention), 128 experts (6 selected + 2 shared), 48 layers, 128K vocab. 6 languages (ko/en/ja/zh/th/vi). 32K native context, 128K with YaRN. Custom tokenizer with 30%+ Korean efficiency gain.

Instruct: MT-Bench 8.42, MMLU 80.80, MATH 86.26, HumanEval+ 79.88. Thinking: MMLU-Pro 75.3, GPQA-Diamond 61.3, AIME 2024 78.3. Multi-turn tool-calling 3x improved over Kanana-1.5-32.5B. Comparable to Qwen3-30B-A3B. CC-BY-NC-4.0.