harrier-oss-v1-0.6b

harrier-oss-v1-0.6b

Turns text in many languages into number vectors for search, clustering and similarity matching, a 0.6B embedding model from Microsoft for developers building retrieval and ranking tools.

Made by microsoft on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: harrier-oss-v1-0.6b is the mid-sized member of Microsoft's harrier-oss-v1 family of multilingual text embedding models. It converts a piece of text into a vector of 1,024 numbers, so that texts with similar meaning end up close together. It has 596M parameters and accepts up to 32,768 tokens.

Who it suits: It suits developers and data teams who need search, matching or grouping of text across many languages without calling an outside service.

Title: Who uses this model
Industry | Who uses it | How they use it | Example
Software | Search and app developers | Rank passages for a search query | Find the passage that best answers a question
Data and analytics | Data teams | Group similar documents | Cluster support tickets by topic
Localization | Language engineers | Match text with its translation | Mine sentence pairs across languages
AI products | Retrieval builders | Feed a retrieval step in a question answering tool | Embed documents once and compare queries to them

If you need a model that writes text, you can skip it, since this one only produces vectors.

Where to use it:

• Python scripts that use the sentence-transformers library
• Python scripts that use the transformers library
• Windows, macOS or Linux with Python

Title: Jobs it does
Use case | Who | What they get back
Retrieval | Search builders | Scores that show which passages match a query
Clustering | Analysts | Vectors you can group by meaning
Semantic similarity | Developers | A closeness score between two texts
Classification and reranking | ML engineers | Vectors that feed a classifier or a reranker

How to set it up: You need Python and a library that can load the model.

• Python with pip
• The sentence-transformers library, or transformers
• About 1.2 GB of disk space for BF16 weights

1. Install sentence-transformers.
2. Optionally save the files first with the two download commands.
3. Run the sample script. It loads the model and encodes two queries and two passages.

pip install -U huggingface_hub
hf download microsoft/harrier-oss-v1-0.6b

How to use it: Encode your queries with a prompt name, encode your documents without one, then multiply the two sets of vectors to get scores. The model's config file defines ready-made prompt names for query types, including web_search_query. You can also pass your own instruction as the prompt.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("microsoft/harrier-oss-v1-0.6b", model_kwargs={"dtype": "auto"})

queries = [
    "how much protein should a female eat",
    "summit define",
]
documents = [
    "As a general guideline, the CDC's average requirement of protein for women ages 19 to 70 is 46 grams per day. But, as you can see from this chart, you'll need to increase that if you're expecting or training for a marathon. Check out the chart below to see how much protein you should be eating each day.",
    "Definition of summit for English Language Learners. : 1  the highest point of a mountain : the top of a mountain. : 2  the highest level. : 3  a meeting or series of meetings between the leaders of two or more governments."
]

query_embeddings = model.encode(queries, prompt_name="web_search_query")
document_embeddings = model.encode(documents)

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

Example: You encode a nutrition question and two passages, one about daily protein for women and one defining a mountain summit. The printed scores are higher for the protein passage, which is the one that answers the question.

For newcomers and experts: Think of an embedding as a fingerprint of a text's meaning. Two texts that mean similar things get fingerprints that sit close together. A first small test is the sample script with your own question and two short passages.

The model is decoder-only. It pools the final token and normalizes the vector with L2. Training used contrastive objectives on a large multilingual mix, and the 270m and 0.6b sizes also learned by distillation from larger embedding models. Queries want a one-sentence task instruction, while documents need none.

Text goes in
Decoder-only model reads it
Final token is pooled
L2-normalized vector comes out

Good to know: Hugging Face lists many languages. The card ranks the family on Multilingual MTEB v2 and places the 0.6b size between the 270m and 27b models. Other people have published 28 quantized and 15 fine-tuned versions.

Title: Multilingual MTEB v2 score by size
270m: 66.5 points
0.6b: 69.0 points
27b: 74.3 points

• License: MIT.
• Cost to run: none named.
• Data it reaches: the files come from Hugging Face, and the text you encode stays on your own computer.
• Not for: writing or chatting, since it only turns text into vectors.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 596M
• Weights: about 1.2 GB in BF16; about 298 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 218,439
• Likes: 338
• License: mit
• Languages: af, am, ar, as, az, be, bg, bn
• Quantized versions: 28
• Fine-tunes: 17
• Tags: sentence-transformers, safetensors, qwen3, feature-extraction, mteb, transformers, multilingual, text-embeddings-inference

Links

• Model: https://huggingface.co/microsoft/harrier-oss-v1-0.6b
• Files: https://huggingface.co/microsoft/harrier-oss-v1-0.6b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/harrier-oss-v1-0.6b

Then follow the usage in its model card. Nothing in this kit runs by itself.