Kanana-2 30B-A3B Instruct
kakaocorp/kanana-2-30b-a3b-instruct is Kakao’s open model aimed at “agentic” workflows: tool calling, multi-step instruction following, and stronger reasoning. Architecturally it’s an MoE model listed as 30B parameters total with ~3B active, and it uses multi-head latent attention (MLA) to keep long-context performance practical.
Two practical details stand out from the card: a native 32,768 token context window, and explicit multilingual coverage (Korean, English, Japanese, Chinese, Thai, and Vietnamese) backed by a newly trained tokenizer. Kakao also calls out a data provenance claim (“no Kakao user data” in pre- or post-training), which matters if you’re evaluating models for enterprise or regulated deployments.
If you try it, start with a function-calling / tool-use harness (even a toy “search + summarize + cite” loop) and measure how often it produces well-formed tool arguments and clean follow-up reasoning. Then scale up to long-context tasks (multi-document QA or long meeting-note summarization) where MLA + a 32k window should make a noticeable difference.
Quick stats from the listing feed: pipeline: text-generation · 26 likes · 2469 downloads.
Source listing: https://huggingface.co/models?sort=modified