🤗 kanana-2-3b-embedding  |  🤗 kanana-2-1.3b-embedding  |  📕 Blog

kanana-2-3b-embedding

News 🔥

• 2026/10/29: 🤗 Released the kanana-2-3b-embedding and kanana-2-1.3b-embedding model weights.
• 2026/07/28: 📕 Published a blog post about the development of the Kanana-2 SLM series.

Introduction

Kanana-2-embedding is Kakao's text embedding model series for Korean and English retrieval. This release includes 3B and 1.3B models. The models are built on Kakao's Kanana-2 SLMs, converted to bidirectional attention with mean pooling. They produce text embeddings for semantic search, retrieval-augmented generation (RAG), and text similarity.

Highlights

• Korean retrieval: kanana-2-3b-embedding achieves the best average score on nano-BEIR (ko) and Ko-Retrieval among the compared models at every Matryoshka dimension.
• Matryoshka representation learning (MRL): Embeddings can be truncated to 1024 / 512 / 256 / 128 dimensions to reduce storage and search cost.
• Long input: The models accept inputs of up to 4,096 tokens.

Model Downloads

Model	Embedding dim	Download
kanana-2-3b-embedding	2560	🤗 HuggingFace
kanana-2-1.3b-embedding	1280	🤗 HuggingFace

Embedding Performance

All models are evaluated with MTEB under the same conditions. Every evaluation runs without any task prompt (raw input text, with no query/document prefix or instruction template). Each cell is the average score of the benchmark, reported per Matryoshka dimension (full / 512 / 256 / 128).

Model	nano-BEIR (en)				nano-BEIR (ko)				MTEB (eng, v2)				Ko-Retrieval†
	full	512	256	128	full	512	256	128	full	512	256	128	full	512	256	128
embeddinggemma-300m	0.5330	0.5059	0.4813	0.4379	0.4564	0.4423	0.4163	0.3764	0.6247	0.6155	0.6097	0.5894	0.7295	0.7172	0.6992	0.6397
llama-nemotron-embed-1b-v2	0.6466	0.6259	0.5961	0.5386	0.5471	0.5204	0.4818	0.4225	0.6369	0.6222	0.6055	0.5752	0.7985	0.7787	0.7506	0.6780
Qwen3-Embedding-0.6B	0.5822	0.5722	0.5472	0.5316	0.4845	0.4712	0.4494	0.4256	0.6311	0.6268	0.6160	0.6013	0.7279	0.7126	0.6825	0.6365
Qwen3-Embedding-4B	0.6379	0.6187	0.5910	0.5449	0.5522	0.5177	0.4912	0.4515	0.6661	0.6606	0.6470	0.6284	0.7837	0.7617	0.7266	0.6854
kanana-nano-2.1b-embedding-v1.0.0‡	0.5285	—	—	—	0.4948	—	—	—	0.6028	—	—	—	0.7579	—	—	—
kanana-essence-8b-embedding-v1.0.0‡	0.6013	—	—	—	0.5518	—	—	—	0.6444	—	—	—	0.7896	—	—	—
kanana-2-1.3b-embedding	0.6332	0.6257	0.6078	0.5603	0.5866	0.5812	0.5610	0.5164	0.6197	0.6037	0.5879	0.5572	0.7958	0.7891	0.7706	0.7211
kanana-2-3b-embedding	0.6552	0.6327	0.6170	0.5899	0.6088	0.5943	0.5683	0.5206	0.6317	0.6203	0.6057	0.5760	0.8211	0.8066	0.7873	0.7488

Per-column best scores are bolded.

† Average over six Korean retrieval tasks: MIRACLRetrieval, Ko-StrategyQA, AutoRAGRetrieval, LawIRKo, SQuADKorV1Retrieval, and PublicHealthQA.

‡ Kanana-1 embedding models do not support Matryoshka truncation, so only the full-dim score is reported.

Library versions used for evaluation: transformers==5.8.0, sentence-transformers==5.5.1, vllm==0.20.2, torch==2.11.0+cu129.

Deployment

The model supports Matryoshka truncation to 1024 / 512 / 256 / 128 dimensions.

sentence-transformers

pip install torch sentence-transformers

from sentence_transformers import SentenceTransformer

model_id = "kakaocorp/kanana-2-3b-embedding"
model = SentenceTransformer(model_id, trust_remote_code=True)

queries = ["What is the capital of France?"]
documents = ["Paris is the capital of France."]

q = model.encode(queries)
d = model.encode(documents)
scores = model.similarity(q, d)
print(scores)

With MRL truncation:

DIM = 256  # one of: 1024, 512, 256, 128

model = SentenceTransformer(model_id, trust_remote_code=True, truncate_dim=DIM)
embeddings = model.encode(["query text", "document text"])

transformers

pip install torch transformers

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

model_id = "kakaocorp/kanana-2-3b-embedding"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
).cuda().eval()

texts = ["query text", "document text"]
inputs = tokenizer(
    texts,
    padding=True,
    truncation=True,
    return_tensors="pt",
    max_length=4096,
).to("cuda")

with torch.no_grad():
    out = model(**inputs)

hidden = out.last_hidden_state
mask = inputs["attention_mask"].unsqueeze(-1).float()
pooled = (hidden * mask).sum(dim=1) / mask.sum(dim=1)
embeddings = F.normalize(pooled, dim=-1)

With MRL truncation:

DIM = 256  # one of: 1024, 512, 256, 128
embeddings = F.normalize(pooled[:, :DIM], dim=-1)

vLLM

pip install vllm

import numpy as np
from vllm import LLM

model_id = "kakaocorp/kanana-2-3b-embedding"

llm = LLM(
    model=model_id,
    runner="pooling",
    convert="embed",
    trust_remote_code=True,
    enable_prefix_caching=False,
    enable_chunked_prefill=False,
)

outputs = llm.embed(["query text", "document text"])
raw = np.asarray([o.outputs.embedding for o in outputs])
embeddings = raw / np.linalg.norm(raw, axis=-1, keepdims=True)

With MRL truncation:

DIM = 256  # one of: 1024, 512, 256, 128
trunc = raw[:, :DIM]
embeddings = trunc / np.linalg.norm(trunc, axis=-1, keepdims=True)

License

The model weights are released under the KananaOpenLicense.

Citation

@misc{kanana2embedding2026,
  title = {{kanana-2-embedding}},
  author = {Doohae Jung},
  year = {2026},
  url = {https://huggingface.co/collections/kakaocorp/kanana-2-slm}
}

Contact

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