Kanana-2 SLM series
modelYour notes
On-device counterpart to the Kanana-2-30B MoE line: a 3B trained from scratch on TPUs and a 1.3B built via cascade pruning + distillation from it, with 3:1 sliding-window hybrid attention (~72.7% KV-cache cut at 32K) and a new tokenizer (+30% Korean efficiency). Kanana Open License (commercial use OK); a 0.9B is mentioned but unreleased.
On 21 October 2026 Kakao released kanana-2-vision-encoder in medium (426M) and tiny (98M) sizes under the Kanana Open License. The model card says both are "based on Kakao's Kanana-2-0.9B LLM", with the text generation module removed. They take images at native resolution and use sliding-window attention and GQA. Kakao compares them against AIMv2, SigLIP2, Perception Encoder, MobileCLIP2 and Qwen3.5's vision encoders over 42 benchmarks in nine categories, all paired with kanana-2-1.3b-instruct under the same training setup. The 0.9B LLM itself is still unreleased.
Model Details
Variants
| Name | Parameters | Notes |
|---|---|---|
| kanana-2-3b | 3B | From scratch on TPUs |
| kanana-2-1.3b | 1.3B | Cascade pruning + distillation from the 3B |