News 🔥

• ✨2025/07/24: Published a blog post about Kanana-1.5-15.7B-A3B models and released 🤗HF model weights.
• 📕2025/05/23: Published a blog post about Kanana 1.5 models and released 🤗HF model weights.
• 📜2025/02/27: Released Technical Report and 🤗HF model weights.
• 📕2025/01/10: Published a blog post about the development of Kanana Nano model.
• 📕2024/11/14: Published blog posts (pre-training, post-training) about the development of Kanana models.
• ▶️2024/11/06: Published a presentation video about the development of the Kanana models.

Table of Contents

• Kanana-1.5-15.7B-A3B
  • Performance
    • Base Model Evaluation
    • Instruct Model Evaluation
• Contributors
• Citation
• Contact

Kanana-1.5-15.7B-A3B

Introducing Kanana-1.5-15.7B-A3B, the first Mixture-of-Experts (MoE) model in our Kanana family, engineered for exceptional efficiency and powerful performance. Kanana-1.5-15.7B-A3B, which has sparse architecture, delivers capabilities comparable to the Kanana-1.5-8B dense model while utilizing only 37% of the FLOPS per token, making it a highly inference-efficient and cost-effective solution for real-world applications. Furthermore, Kanana-1.5-15.7B-A3B is powered by our newly enhanced post-training strategy, which includes on-policy distillation followed by reinforcement learning.

[!Note] Neither the pre-training nor the post-training data includes Kakao user data.

Performance

Base Model Evaluation

Instruct Model Evaluation

[!Note] * This model is not an open-sourced, just for comparison with Kanana-1.5-15.7B-A3B

Evaluation Protocol

• Base Model Benchmarks

  • MMLU, KMMLU, HAE-RAE: 5-shot, log-likelihood
  • HumanEval: 0-shot, pass@1
  • MBPP: 3-shot, pass@1
  • GSM8K: 5-shot, exact-match (strict-match)

• Instruct Model Benchmarks

  • MT-Bench, KoMT-Bench: 0-shot, gpt-4o-2024-08-06 as judge model
  • IFEval: 0-shot, mean of strict-prompt-level and strict-instruction-level
  • HumanEval+, MBPP+: 0-shot, pass@1
  • GSM8K, MATH: 0-shot, rule-based verification

Quickstart

vLLM

• vllm>=0.8.5 or the latest version is required to run Kanana model.

Example Usage for Kanana-1.5-15.7B-A3B-Base

vllm serve $path_to_model \
        --served_model_name kanana-1.5-15.7b-a3b-base \
        --max-model-len 32768 \
        --gpu-memory-utilization 0.9 \
        --port 8000 \
        --dtype auto \
        --disable_cascade_attn

curl http://localhost:8000/v1/completions -H "Content-Type: application/json" -d '{
    "model": "kanana-1.5-15.7b-a3b-base",
    "prompt": "Kakao is a leading company in South Korea, and it is known for ",
    "max_tokens": 32,
    "top_k": 1
}'

# Output:
'''
...
"choices":[{"index":0,"text":"1) its innovative technology, 2) its high-quality products, and 3) its strong brand image. The company has a long history of success,"...
...
'''

Contributors

• Language Model Training
  • Yunju Bak, Doohae Jung, Boseop Kim, Nayeon Kim, Hojin Lee, Jaesun Park, Minho Ryu, Jiyeon Ham, Seungjae Jung, Hyunho Kim, Hyunwoong Ko, Changmin Lee, Taegyeong Eo

Citation

@misc{kananallmteam2025kananacomputeefficientbilinguallanguage,
      title={Kanana: Compute-efficient Bilingual Language Models}, 
      author={Kanana LLM Team and Yunju Bak and Hojin Lee and Minho Ryu and Jiyeon Ham and Seungjae Jung and Daniel Wontae Nam and Taegyeong Eo and Donghun Lee and Doohae Jung and Boseop Kim and Nayeon Kim and Jaesun Park and Hyunho Kim and Hyunwoong Ko and Changmin Lee and Kyoung-Woon On and Seulye Baeg and Junrae Cho and Sunghee Jung and Jieun Kang and EungGyun Kim and Eunhwa Kim and Byeongil Ko and Daniel Lee and Minchul Lee and Miok Lee and Shinbok Lee and Gaeun Seo},
      year={2025},
      eprint={2502.18934},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18934}, 
}

Contact

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