Harrier Embedding 0.6B is a highly efficient multilingual text embedding model developed by Microsoft, built on top of the Qwen3 architecture. It is designed to generate fixed-size sentence embeddings suitable for semantic similarity, dense retrieval, and clustering tasks across 94+ languages.
Despite its compact size of just few million parameters, it achieves strong performance on embedding benchmarks, scoring 69.0 on Multilingual MTEB v2, making it an excellent choice for resource-constrained environments requiring high-quality multilingual retrieval.
Architecture Highlights:
Base Architecture: Qwen3 (Transformer Decoder)
Pooling Strategy: Last non-padding token pooling (default)
Normalization: L2 normalized embeddings (default)
Languages Supported: 94+
Intended Use
Semantic Search / RAG: Generating embeddings for documents and queries in Retrieval-Augmented Generation systems.
Clustering: Grouping similar text documents together based on spatial distance.
Classification: Using the resulting embeddings as features for downstream classification tasks.
Training
All models are trained with contrastive learning objectives on a large-scale mixture of multilingual datasets covering diverse tasks. The 270m and 0.6b variants are additionally trained with knowledge distillation from larger embedding models.
Model Summary
A 270M parameter multilingual text embedding model based on the Gemma 3 architecture by Microsoft. Fine-tuned for dense retrieval and semantic similarity across 94+ languages, achieving 66.5 on Multilingual MTEB v2.
Overview
Harrier-OSS 270M is a highly efficient, multilingual text embedding model developed by Microsoft. It is built upon the decoder-only Gemma 3 270M architecture and fine-tuned specifically for dense vector representations.
The model maps variable-length text inputs into fixed-size 640-dimensional vectors, making it ideal for downstream NLP tasks such as:
Semantic Search and Dense Retrieval
Clustering and Topic Modeling
Retrieval-Augmented Generation (RAG) pipelines
Zero-shot cross-lingual information retrieval
Despite its small size (268M parameters), Harrier-OSS 270M achieves highly competitive performance across 94+ languages, scoring a 66.5 on the Multilingual MTEB v2 benchmark.
Architecture
Backbone: Gemma 3 (Decoder-only causal language model)
Parameters: 268 Million
Hidden Dimension / Output Size: 640
Pooling Strategy: Last-token pooling (standard for causal architectures)
Normalization: L2-normalized outputs (ready for fast dot-product similarity)
pip install -U -q keras-hub
pip install -U -q keras
Jax, TensorFlow, and Torch come preinstalled in Kaggle Notebooks. For instructions on installing them in another environment see the Keras Getting Started page.
Presets
Preset Name
Parameters
Description
harrier_embedding_oss_0.6B
0.6B
Multilingual text embedding model based on the Qwen3-0.6B architecture, fine-tuned for dense retrieval and semantic similarity across 94+ languages.
harrier_embedding_oss_270m
270M
Multilingual text embedding model based on the Gemma3-270M architecture, fine-tuned for dense retrieval and semantic similarity across 94+ languages.
Example Usage
Basic Semantic Search
import keras_hub
# Load the embedder
embedder = keras_hub.models.Qwen3TextEmbedder.from_preset(
"harrier_embedding_oss_0.6b",
)
# Define query and documents
query = "Which planet is known as the Red Planet?"
documents = [
"Mars is often referred to as the Red Planet.",
"Venus is often called Earth's twin.",
]
# For best results with Harrier, prepend the instruction to queries
q_emb = embedder.encode_text(
"Instruct: Given a web search query, retrieve relevant passages\n"
f"Query: {query}"
)
d_embs = embedder.encode_text(documents)
# Compute cosine similarity
sims = embedder.similarity(q_emb, d_embs)
best_match_idx = sims.numpy().argmax()
print(f"Best match: {documents[best_match_idx]}")
Extracting Raw Embeddings
import keras_hub
embedder = keras_hub.models.Qwen3TextEmbedder.from_preset(
"harrier_embedding_oss_0.6b",
)
# Encode a batch of texts
embeddings = embedder.encode_text([
"The quick brown fox jumps over the lazy dog.",
"Keras makes machine learning accessible."
])
print(embeddings.shape) # (2, 1024)
Example Usage with Hugging Face URI
Basic Semantic Search
import keras_hub
# Load the embedder
embedder = keras_hub.models.Qwen3TextEmbedder.from_preset(
"hf://keras/harrier_embedding_oss_0.6b",
)
# Define query and documents
query = "Which planet is known as the Red Planet?"
documents = [
"Mars is often referred to as the Red Planet.",
"Venus is often called Earth's twin.",
]
# For best results with Harrier, prepend the instruction to queries
q_emb = embedder.encode_text(
"Instruct: Given a web search query, retrieve relevant passages\n"
f"Query: {query}"
)
d_embs = embedder.encode_text(documents)
# Compute cosine similarity
sims = embedder.similarity(q_emb, d_embs)
best_match_idx = sims.numpy().argmax()
print(f"Best match: {documents[best_match_idx]}")
Extracting Raw Embeddings
import keras_hub
embedder = keras_hub.models.Qwen3TextEmbedder.from_preset(
"hf://keras/harrier_embedding_oss_0.6b",
)
# Encode a batch of texts
embeddings = embedder.encode_text([
"The quick brown fox jumps over the lazy dog.",
"Keras makes machine learning accessible."
])
print(embeddings.shape) # (2, 1024)
