Kanana-2
Open-source multilingual model for agentic AI and reasoning.
Kakao Corp released Kanana-2, an open-weight language model designed for agentic tasks and complex reasoning. It uses a mixture of experts architecture to balance performance with efficiency across six languages.
- Uses 3 billion active parameters out of 30 billion total.
- Supports Korean, English, Japanese, Chinese, Thai, and Vietnamese.
- Native context window of 32,768 tokens, extendable to 128K.
- Designed for agentic AI with enhanced tool calling capabilities.
- No Kakao user data was used for training the model.
What it is
Kanana-2 is a large language model developed by Kakao Corp. It belongs to the Kanana family and focuses on agentic artificial intelligence. The model supports tool calling, follows complex instructions, and performs logical reasoning. It features a total of 30 billion parameters but activates only 3 billion during operation. This design aims to provide high throughput while maintaining strong performance. The system handles text in six languages and supports context windows of up to 32,768 tokens natively.
How it came about
This model represents the latest evolution in the Kanana series. It succeeds a previous version that contained 32.5 billion parameters. The developers state that no user data from Kakao services was used during pre-training or post-training phases. The release includes base, instruct, and thinking variants available on Hugging Face. The team introduced a new tokenizer to improve efficiency, particularly for Korean text.
Inside the model
The architecture employs a Mixture of Experts approach with Multi-head Latent Attention. It selects six experts for each token processing step. This structure allows the model to use fewer active parameters than its predecessor. A newly trained tokenizer improves compression rates by over thirty percent for Korean. The default context length is 32,768 tokens. Users can extend this to 128,000 tokens by applying YaRN scaling techniques via configuration changes.
What it can do
Kanana-2 is built for agentic workflows, including tool use and multi-turn interactions. It includes specific reasoning models designed for deliberate thinking on hard problems. Evaluations were conducted using internal toolkits and external judges like GPT-4o. The model supports Korean, English, Japanese, Chinese, Thai, and Vietnamese. Performance metrics cover multiple-choice questions, mathematics, and live benchmarks.
How to use it
The model weights are open and available under the Kanana License. Developers can download base, instruct, and thinking versions from Hugging Face. Deployment is supported through libraries such as vLLM and SGLang. These tools allow the creation of OpenAI-compatible API endpoints. The code requires transformers library version 4.51.0 or higher for proper execution.
Safety and testing
The developers explicitly state that no Kakao user data was utilized for training. The model includes specific variants for reasoning and instruction following. Internal evaluations were performed to assess performance on various benchmarks. No external safety audits or red-teaming results are detailed in the provided documentation.
Questions people ask about Kanana-2
What languages does Kanana-2 support?
The model supports six languages: Korean, English, Japanese, Chinese, Thai, and Vietnamese. It uses a new tokenizer that improves efficiency for these languages, with a notable thirty percent improvement for Korean text processing.
How many parameters does the model have?
Kanana-2 has a total of 30 billion parameters. However, it uses a Mixture of Experts architecture that activates only 3 billion parameters per token. This design helps maintain high throughput and efficiency during inference.
Can I use Kanana-2 for commercial projects?
The model weights are released under the Kanana License. You should review the specific terms of this license to determine if it permits your intended commercial use. The weights are openly available for download on Hugging Face.
What is the maximum context length?
The model natively supports a context length of 32,768 tokens. Users can extend this to 128,000 tokens by applying YaRN scaling methods. This requires updating the configuration file with specific rope scaling parameters.
Good reading and sources
There is no English Wikipedia article on Kanana-2 yet, so the facts on this page come from the sources listed above: huggingface.co/kakaocorp/kanana-2-30b-a3b-instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.