Skip to content
← Models
Open weightsKakao/10 trustMetrics-checked

Kanana-2 3B

Kakao's on-device Korean-English small language model, open sourced.

Visit official page →

About

Kanana-2 3B is part of Kakao's Kanana-2 series of small language models, open-sourced on Hugging Face in 2026 in base and instruct variants at 1.3B and 3B sizes. The models are optimized for on-device deployment on smartphones, with a Korean-optimized tokenizer that Kakao says improves Korean processing efficiency by over 30 percent versus prior versions. Kakao reports performance on par with global models of the same class on conversation, knowledge, code, math, instruction following, and tool calling.

The Kanana-2 instruct models also went through Kakao's AI Safety Evaluation Platform and ranked first among similar-size global models on the Korean AssurAI safety benchmark. Release date is era-level (July 2026) since reporting gives the announcement date but not an exact day. License is the custom Kanana license, so commercial terms should be checked before production use.

Specs

Context window

Pricing

Free quota

API access

Weights

Open weights

Released

July 28, 2026

Benchmarks

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

on-device-aikorean-chatmobile-assistants

Trust score — 7/10

/10 trustMetrics-checked

Scored 7/10 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 115; Trending score 72/100 from researcher assessment.

More in the kanana family

Sources