We introduce Kanana, a series of bilingual language models (developed by Kakao) that demonstrate exceeding performance in Korean and competitive performance in English. The computational cost of Kanana is significantly lower than that of state-of-the-art models of similar size. The report details the techniques employed during pre-training to achieve compute-efficient yet competitive models, including high-quality data filtering, staged pre-training, depth up-scaling, and pruning and distillation. Furthermore, the report outlines the methodologies utilized during the post-training of the Kanana models, encompassing supervised fine-tuning and preference optimization, aimed at enhancing their capability for seamless interaction with users. Lastly, the report elaborates on plausible approaches used for language model adaptation to specific scenarios, such as embedding, function calling, and Retrieval Augmented Generation (RAG). The Kanana model series spans from 2.1B to 32.5B parameters with 2.1B models (base, instruct, embedding, function call, and RAG) publicly released to promote research on Korean language models.

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

News
Performance
Quickstart
License
Citation
Contributors
Contact

News

📜2025/02/27: Released Technical Report and 🤗HF model weights.
📕2025/01/10: Published a blog post about the development of Kanana-Nano model. (Kanana-Nano)
📕2024/11/14: Published blog posts about the development of Kanana models. (Kanana LLM: Pre-training, Kanana LLM: Post-training)
▶️2024/11/06: Published a presentation video about the development of the Kanana models. (if(kakaoAI)2024)

Performance

Below are partial report on the performance of the Kanana model series. Please refer to the Technical Report for the full results.

Pre-trained Model Performance

Models	MMLU	KMMLU	HAERAE	HumanEval	MBPP	GSM8K
27b+ scale
Kanana-Flag-32.5b	77.68	62.10	90.47	51.22	63.40	70.05
Qwen2.5-32b	83.10	63.15	75.16	50.00	73.40	82.41
Gemma-2-27b	75.45	51.16	69.11	51.22	64.60	74.37
EXAONE-3.5-32b	72.68	46.36	82.22	-	-	-
Aya-Expanse-32b	74.52	49.57	80.66	-	-	-
7b+ scale
Kanana-Essence-9.8b	67.61	50.57	84.98	40.24	53.60	63.61
Llama-3.1-8b	65.18	41.02	61.78	35.37	48.60	50.87
Qwen2.5-7b	74.19	51.68	67.46	56.71	63.20	83.85
Gemma-2-9b	70.34	48.18	66.18	37.20	53.60	68.16
EXAONE-3.5-7.8b	65.36	45.30	77.54	-	-	-
Aya-Expanse-8b	62.52	40.11	71.95	-	-	-
2b+ scale
Kanana-Nano-2.1b	54.83	44.80	77.09	31.10	46.20	46.32
Llama-3.2-3b	56.40	35.57	47.66	25.61	39.00	27.37
Qwen2.5-3b	65.57	45.28	61.32	37.80	55.60	69.07
Gemma-2-2b	52.89	30.67	45.55	20.12	28.20	24.72
EXAONE-3.5-2.4b	59.27	43.58	69.65	-	-	-
70b+ scale
Llama-3.1-70b	78.93	53.00	76.35	57.32	66.60	81.73
Qwen2.5-72b	86.12	68.57	80.84	55.49	76.40	92.04

Post-trained Model Performance

Instruction-following Benchmarks

Models	MT-Bench	LogicKor	KoMT-Bench	WildBench	IFEval
27b+ scale
Kanana-Flag-32.5b	8.356	9.524	8.058	54.14	0.856
Qwen2.5-32b	8.331	8.988	7.847	51.13	0.822
Gemma-2-27b	8.088	8.869	7.373	46.46	0.817
EXAONE-3.5-32b	8.375	9.202	7.907	54.30	0.845
Aya-Expanse-32b	7.788	8.941	7.626	48.36	0.735
7b+ scale
Kanana-Essence-9.8b	7.769	8.964	7.706	47.27	0.799
Llama-3.1-8b	7.500	6.512	5.336	33.20	0.772
Qwen2.5-7b	7.625	7.952	6.808	41.31	0.760
Gemma-2-9b	7.633	8.643	7.029	40.92	0.750
EXAONE-3.5-7.8b	8.213	9.357	8.013	50.98	0.826
Aya-Expanse-8b	7.131	8.357	7.006	38.50	0.645
2b+ scale
Kanana-Nano-2.1b	6.400	7.964	5.857	25.41	0.720
Llama-3.2-3b	7.050	4.452	3.967	21.91	0.767
Qwen2.5-3b	6.969	6.488	5.274	25.76	0.355
Gemma-2-2b	7.225	5.917	4.835	28.71	0.428
EXAONE-3.5-2.4b	7.919	8.941	7.223	41.68	0.790
70b+ scale
Llama-3.1-70b	8.275	8.250	6.970	46.50	0.875
Qwen2.5-72b	8.619	9.214	8.281	55.25	0.861

General Benchmarks

Models	MMLU	KMMLU	HAE-RAE	HumanEval+	MBPP+	GSM8K	MATH
27b+ scale
Kanana-Flag-32.5b	81.08	64.19	68.18	77.44	69.84	90.83	57.82
Qwen2.5-32b	84.40	59.37	48.30	82.32	71.96	95.30	81.90
Gemma-2-27b	78.01	49.98	46.02	70.12	70.90	91.05	53.80
EXAONE-3.5-32b	78.30	55.44	52.27	78.66	70.90	93.56	76.80
Aya-Expanse-32b	74.49	42.35	51.14	64.63	65.61	75.06	42.82
7b+ scale
Kanana-Essence-9.8b	70.64	50.76	47.16	72.56	69.05	84.91	42.24
Llama-3.1-8b	71.18	39.24	40.91	60.98	57.67	82.71	49.86
Qwen2.5-7b	77.23	46.87	37.50	73.78	70.63	91.58	75.22
Gemma-2-9b	73.47	44.47	39.77	59.76	64.55	87.72	48.10
EXAONE-3.5-7.8b	72.62	52.09	46.02	79.27	66.67	89.99	73.50
Aya-Expanse-8b	61.23	35.78	39.20	42.68	56.88	78.85	30.80
2b+ scale
Kanana-Nano-2.1b	52.48	38.51	33.52	63.41	62.43	72.32	29.26
Llama-3.2-3b	56.09	3.07	17.05	56.71	50.26	66.57	38.18
Qwen2.5-3b	69.18	38.33	32.39	67.68	64.02	84.00	65.72
Gemma-2-2b	57.69	6.99	7.95	35.37	45.24	49.81	21.68
EXAONE-3.5-2.4b	63.19	14.27	14.20	70.73	59.79	83.78	64.04
70b+ scale
Llama-3.1-70b	83.48	39.08	53.41	75.61	66.40	91.66	63.98
Qwen2.5-72b	87.14	65.78	60.80	81.10	75.66	95.45	82.60

Embedding Model Performance

Backbone	Kanana-Nano-2.1b	Llama-3.2-3b	Qwen2.5-3b	Llama-3.2-1b	Qwen-2.5-1.5b
English	51.56	53.28	54.00	48.77	50.60
Korean	65.00	59.43	62.10	54.68	54.60
Avg.	58.28	56.35	58.05	51.73	52.60

Quickstart

Dependencies

The following versions have been tested:

bash

# transformers
pip install transformers==4.57.6

# sentence-transformers
pip install sentence-transformers==5.2.2

# vllm
pip install vllm==0.15.1

# common
pip install datasets torch tqdm numpy

Tested with: transformers==4.57.6, sentence-transformers==5.2.2, vllm==0.15.1, torch==2.9.1+cu129.

🤗 HuggingFace Transformers

[!Note] trust_remote_code=True is required. You also need datasets (pip install datasets) for the _do_encode() batched method.

python

import torch.nn.functional as F
from transformers import AutoModel

instruction = "Given a question, retrieve passages that answer the question"
queries = [
    "are judo throws allowed in wrestling?",
    "how to become a radiology technician in michigan?",
]

passages = [
    "Since you're reading this, you are probably someone from a judo background or someone who is just wondering how judo techniques can be applied under wrestling rules. So without further ado, let's get to the question. Are Judo throws allowed in wrestling? Yes, judo throws are allowed in freestyle and folkstyle wrestling. You only need to be careful to follow the slam rules when executing judo throws. In wrestling, a slam is lifting and returning an opponent to the mat with unnecessary force.",
    "Below are the basic steps to becoming a radiologic technologist in Michigan:Earn a high school diploma. As with most careers in health care, a high school education is the first step to finding entry-level employment. Taking classes in math and science, such as anatomy, biology, chemistry, physiology, and physics, can help prepare students for their college studies and future careers.Earn an associate degree. Entry-level radiologic positions typically require at least an Associate of Applied Science. Before enrolling in one of these degree programs, students should make sure it has been properly accredited by the Joint Review Committee on Education in Radiologic Technology (JRCERT).Get licensed or certified in the state of Michigan.",
]

model = AutoModel.from_pretrained(
    "kakaocorp/kanana-nano-2.1b-embedding",
    trust_remote_code=True,
).to("cuda")

max_length = 512
query_embeddings = model.encode(queries, instruction=instruction, max_length=max_length)
passage_embeddings = model.encode(passages, instruction="", max_length=max_length)

# get the embeddings with DataLoader (spliting the datasets into multiple mini-batches)
# batch_size = 2
# query_embeddings = model._do_encode(queries, batch_size=batch_size, instruction=instruction, max_length=max_length)
# passage_embeddings = model._do_encode(passages, batch_size=batch_size, instruction="", max_length=max_length)

query_embeddings = F.normalize(query_embeddings, p=2, dim=1)
passage_embeddings = F.normalize(passage_embeddings, p=2, dim=1)

scores = (query_embeddings @ passage_embeddings.T) * 100
print(scores.tolist())

# Output (bf16, transformers==4.57.6):
# [[84.2861099243164, 31.624895095825195], [35.9390869140625, 81.85345458984375]]

sentence-transformers

The model is compatible with sentence-transformers. The default prompt "query" is automatically applied (configured in config_sentence_transformers.json). Instruction tokens are excluded from mean pooling via include_prompt: false.

python

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("kakaocorp/kanana-nano-2.1b-embedding", trust_remote_code=True)

# Default prompt "query" is applied automatically
query_embeddings = model.encode(queries)

# Passages use no prompt
passage_embeddings = model.encode(passages, prompt_name="document")

AutoModel .encode() and sentence-transformers produce identical results at float32.

vLLM

Kanana2VecModel is not in vLLM's native architecture registry, but the model is compatible via hf_overrides using the built-in LlamaBidirectionalModel (same bidirectional attention and weight layout):

python

from vllm import LLM, EngineArgs

args = EngineArgs(
    model="kakaocorp/kanana-nano-2.1b-embedding",
    runner="pooling",
    convert="embed",
    trust_remote_code=True,
    dtype="auto",
    enforce_eager=True,
    max_model_len=512,
    hf_overrides={
        "architectures": ["LlamaBidirectionalModel"],
        "pooling": "avg",
    },
)
llm = LLM(**vars(args))

instruction = "Instruct: Given a question, retrieve passages that answer the question\nQuery: "
prompts = [instruction + q for q in queries]
outputs = llm.encode(prompts, pooling_task="embed", truncate_prompt_tokens=-1)

[!Note] vLLM's mean pooling operates over all tokens including instruction tokens (it does not support include_prompt: false / pool_mask). The bidirectional attention is functionally identical (cosine >0.9999 with same pooling scope), but the different pooling scope causes cosine similarity of ~0.91–0.93 when compared against HF .encode() or sentence-transformers results.

License

The Kanana models are licensed under CC-BY-NC-4.0.

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}, 
}

Contributors

Pre-training: Yunju Bak, Doohae Jung, Boseop Kim, Nayeon Kim, Hojin Lee, Jaesun Park, Minho Ryu
Post-training: Jiyeon Ham, Seungjae Jung, Hyunho Kim, Hyunwoong Ko, Changmin Lee, Daniel Wontae Nam, Kyoung-Woon On
Adaptation: Seulye Baeg, Junrae Cho, Taegyeong Eo, Sunghee Jung, Jieun Kang, EungGyun Kim, Eunhwa Kim, Byeongil Ko, Daniel Lee, Donghun Lee, Minchul Lee, Miok Lee, Shinbok Lee, Minho Ryu, Gaeun Seo

Contact

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

kakaocorp/kanana-nano-2.1b-embedding

Kanana

Introduction

Table of Contents

News

Performance

Pre-trained Model Performance

Post-trained Model Performance

Instruction-following Benchmarks

General Benchmarks

Embedding Model Performance

Quickstart

Dependencies

🤗 HuggingFace Transformers

sentence-transformers

vLLM

License

Citation

Contributors

Contact

kakaocorp/
kanana-nano-2.1b-embedding