AI Briefing#kanana
The latest AI and developer news about #kanana, with the original source and a short summary.
The latest AI and developer news about #kanana, with the original source and a short summary.
Kakao will terminate the beta service of its standalone AI app 'Kanana' on October 15, 2026, and integrate its features into existing platforms such as KakaoTalk.
The Kanana model achieved superior performance with 11T tokens, half the amount of global targets, and Kakao is reviewing industry-academia collaborations including GPU resource support.
The program expands to three categories—AI experts, creators, and college students—with a five-month activity period and AI usage opportunities worth 1 million won.
The first development log for Kanana-2 summarizes the key decisions made during the pre-training stage.
The Kakao Kanana organization has released 'Kanana-v-embedding', a 2-billion parameter multimodal embedding model, demonstrating superior search performance compared to Jina Embeddings v4 through training on the Korean-specific dataset KoEmbed.
Kakao and the Korean Institute of Information Scientists and Engineers (KIISE) are hosting an AI agent competition with a total prize pool of 30 million KRW, offering bonus points for using the Kanana model or playMCP.
At if(kakao)25, Kakao unveiled a technology evolution roadmap toward Agentic AI, including Kanana-2, which improves inference throughput by applying MLA and MoE, and the omni model Kanana-o, which delivers performance comparable to GPT-4o.
The conference will be held at the Kakao AI Campus in Yongin from September 23 to 25, with registration opening on August 28.
AIME benchmark performance was improved through high learning rates in the SFT stage and relaxed clipping strategies in RL.
By applying the Kanana model, it achieves a 98% receipt recognition rate and automatically recommends settlement categories.
Kakao has released the lightweight multimodal language model 'Kanana-1.5-v-3b' as open source for commercial use, achieving superior performance compared to competing models on Korean benchmarks.
Kakao has open-sourced 'Kanana-1.5-15.7B-A3B', a model utilizing an MoE structure to achieve performance comparable to Dense 8B models with only 3B active parameters.
Kakao has released Kanana 1.5 8B and 2.1B models under the Apache 2.0 license, allowing for commercial use.
Kakao has open-sourced the 'Kanana 1.5' 8B and 2.1B models under the Apache 2.0 license, enhancing math and coding performance and long-context understanding to support Agentic AI.
Kanana Flag demonstrated overwhelming performance on Korean benchmarks while reducing training costs by over 50% compared to third-party models.
The reviews capture the participation experiences of students from Kakao Tech Bootcamp and Gurumtoon Unib in technical sessions covering Kanana, Code Buddy, and more.
Kakao has unveiled the 'Kanana' family of AI models specialized for Korean language processing, demonstrating overwhelming performance in Korean compared to global SOTA models and already applied to services such as KakaoTalk.
Kakao's Kanana Essence, trained on 3T tokens, achieved higher benchmark averages than llama-3.1-8b, while Kanana Nano maximized efficiency through pruning and distillation techniques.
The Kanana LLM, developed through SFT, DPO, and model merging, demonstrated superior performance compared to similar-sized models on Korean benchmarks such as LogicKor.
Implemented natural likeness using an HG-DPO-based backbone model and the Instantbooth module, validated via a photo booth at if(kakaoAI)2024.