Kanana-2 Highlights

Kanana-2, the latest open-source evolution of the Kanana model family, is designed specifically for Agentic AI, presenting substantial enhancements in tool calling, complex instruction following, and logical reasoning. This new version adopts a cutting-edge architecture featuring MLA (Multi-head Latent Attention) and MoE (Mixture of Experts). These innovations allow the model to utilize significantly fewer active parameters compared to the previous 32.5B model while delivering superior performance and ensuring high throughput. Furthermore, the model natively supports context lengths of up to 32,768 tokens, enabling it to maintain coherence when handling extensive documents or long-context interactions.

In addition, Kanana-2 now supports 6 languages, covering Korean, English, Japanese, Chinese, Thai, and Vietnamese. To support this expansion, Kanana-2 utilizes a newly trained tokenizer that demonstrates superior tokenization efficiency across these languages, including an improvement of over 30% specifically for Korean. Finally, to address advanced problem-solving needs, Kanana-2 introduces reasoning models capable of deliberate thinking and reasoning, achieving significantly enhanced performance in downstream tasks, especially when tackling hard problems.

[!NOTE] No Kakao user data was used for either pre-training or post-training.

Model Overview

kanana-2-30b-a3b series has the following features:

• Total Parameters: 30B
• Activated Parameters: 3B
• Number of Layers: 48
• Number of Dense Layers: 1
• Number of Experts: 128
• Number of Selected Experts: 6
• Number of Shared Experts: 2
• Attention Mechanism: MLA
• Vocabulary Size: 128256
• Context Length: 32,768

Model Downloads

Performance

Base model evaluation results

Instruct model evaluation results

Reasoning model evaluation results

Quickstart

You should install the transformers library with version >= 4.51.0.

The following contains a code snippet illustrating how to use the model generate content based on given inputs.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "kakaocorp/kanana-2-30b-a3b-instruct"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="bfloat16",
    device_map="auto"
)

prompt = "Explain the future of AI."

messages = [
    {"role": "user", "content": prompt}
]
input_ids = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

output = model.generate(
    input_ids,
    max_new_tokens=128,
    do_sample=False,
)
print(tokenizer.decode(output[0]))

For deployment, you can use vllm or sglang to create an OpenAI-compatible API endpoint:

• vLLM:

  vllm serve kakaocorp/kanana-2-30b-a3b-instruct --enable-auto-tool-choice --tool-call-parser hermes
  

• SGLang:

  python3 -m sglang.launch_server --model-path kakaocorp/kanana-2-30b-a3b-instruct --tool-call-parser qwen
  

Processing 32K+ Length

Currently, the config.json uploaded to HuggingFace is configured for token lengths of 32,768 or less. To process tokens beyond this length, YaRN must be applied. By updating the config.json with the following parameters, you can apply YaRN to handle token sequences up to 128K in length:

"rope_scaling": {
    "beta_fast": 32,
    "beta_slow": 1,
    "factor": 4.0,
    "mscale": 1.0,
    "mscale_all_dim": 1.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn",
},

Passing command line arguments for deployment:

• vllm

  vllm serve ... --hf-overrides '{"max_position_embeddings": 131072, "rope_scaling": {"rope_type":"deepseek_yarn","factor":4.0,"beta_fast":32,"beta_slow":1,"mscale":1.0,"mscale_all_dim":1.0,"original_max_position_embeddings":32768}}'
  

• sglang

  python3 -m sglang.launch_server ... --json-model-override-args '{"max_position_embeddings":131072, "rope_scaling":{"rope_type":"deepseek_yarn","factor":4.0,"beta_fast":32,"beta_slow":1,"mscale":1.0,"mscale_all_dim":1.0,"original_max_position_embeddings":32768}}'
  

[!NOTE] Most leading open-source implementations of static YaRN apply a constant scaling factor, which can negatively impact performance on shorter texts. To ensure optimal performance:

• Enable rope_scaling only when necessary for processing long contexts.
• Adjust the factor based on your specific needs (e.g., set factor to 2.0 for a 65,536-token context)."

License

The model weights are released under the Kanana License.

Citation

@article{,
  title={Kanana-2 LLM},
  author={Kanana LLM},
  year={2025},
  url={https://huggingface.co/collections/kakaocorp/kanana-2}
}

Contact

• Kanana LLM Team Technical Support: kanana-llm@kakaocorp.com
• Business & Partnership Contact: alpha.k@kakaocorp.com