Kakao has improved its Kanana-2 AI model and released 4 new versions for free use. After showing good performance with the original Kanana-2 model in December, Kakao made a big update. The update adds four new models: Base, Instruct, Thinking (for reasoning), and Mid-training (for research). These new models are more efficient and cost-effective, especially for AI that needs to perform real-world tasks.
The new models are designed to run smoothly on Graphics Processing Units (GPUs) like Nvidia’s A100, which are general-purpose graphics cards. This optimization makes the models more affordable for small businesses and researchers, allowing them to run advanced AI systems without needing expensive, specialized infrastructure.
Kanana-2 has 32 billion parameters, but only 3 billion are used during processing, which helps make it faster and more efficient. The model also has a new training system that helps it learn better by avoiding forgetting old information when it learns new things.
Unlike regular chatbots, the updated Kanana-2 is made for AI that can handle real-world tasks, like following complex instructions and choosing the right tools to use.
Executive Statement
According to Kim Byung-hak, performance lead for Kakao’s Kanana project, the updated Kanana-2 is the result of their deep focus on how to build practical agentic AI without relying on expensive infrastructure. By open-sourcing models that deliver high efficiency even on general-purpose infrastructure, they hope to offer a new alternative for AI adoption and help advance Korea’s AI R&D ecosystem.