harrier-oss-v1-27b

harrier-oss-v1-27b

Turns text in many languages into embedding vectors for search, clustering, similarity, classification and reranking. Made for developers who run semantic search or retrieval and can host a 27B model.

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-v1-27b is the largest model in Microsoft's multilingual text embedding family. It turns a sentence or passage into a dense vector, using a decoder-only design with last-token pooling and L2 normalization. You compare the vectors to find text with a similar meaning. It was trained with contrastive learning on a large mix of multilingual data.

Who it suits: It suits developers and data teams who build search, recommendation or text-grouping features across several languages and have the hardware for a 27B model.

Industry | Who uses it | How they use it | Example
Software | Search engineers | Embed queries and passages, then rank passages by score | A semantic search over help articles
Data science | Analysts | Cluster texts or measure how close two texts are in meaning | Grouping survey answers by topic
Localization | Language teams | Mine matching sentence pairs across languages | Finding translated sentence pairs
Support | Platform teams | Retrieve relevant passages and rerank them | Matching a question to stored answers

Skip it if you need a model that writes replies, since it only outputs vectors. Skip the 27B size if memory is tight, because the family also has 270m and 0.6b models.

Where to use it:

• Python with the sentence-transformers library
• Python with the transformers library
• Windows, macOS and Linux, on a machine that can hold about 54 GB of BF16 weights

Use case | Who | What they get back
Retrieval | Search engineers | Similarity scores between a query and passages
Clustering | Data teams | Vectors you can group by meaning
Bitext mining | Language teams | Vectors that line up matching sentences across languages
Classification | ML engineers | Vectors you feed to a classifier

How to set it up:

• Python, with sentence-transformers or transformers installed
• About 54 GB of disk space for the BF16 weights, or about 13.5 GB at 4-bit as an estimate
• Enough GPU or system memory to run a 27B model

1. Install the Hugging Face tools.
2. Download the model files, or let the library fetch them on first use.
3. Load the model with sentence-transformers and encode your text.

pip install -U huggingface_hub
hf download microsoft/harrier-oss-v1-27b

How to use it: You encode queries with a prompt and documents without one, then take the dot product of the two sets of vectors. The card's example multiplies the result by 100 to print scores. A config file in the repository holds ready-made prompts for web search, similarity and bitext work.

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("microsoft/harrier-oss-v1-27b", model_kwargs={"dtype": "auto"})

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())

To use your own instruction, pass it directly.

model.encode(queries, prompt="Instruct: Retrieve semantically similar text\nQuery: ")

Tool | What it does
SentenceTransformer | Loads the model with the dtype set automatically
encode | Turns a list of texts into embedding vectors
prompt_name | Picks a pre-set instruction such as web_search_query
last_token_pool | Pools the last token's state when you use plain transformers

Example: You encode two questions, "how much protein should a female eat" and "summit define", plus two passages that answer them. The script prints a grid of scores, one row per question. The matching passage should score highest in each row.

For newcomers and experts: An embedding turns text into a list of numbers so that texts with similar meaning land close together. A good first try is the card's two-question script, and reading which passage scores higher for each question.

If you know the field, this is a decoder-only embedder with last-token pooling and L2 normalization. With plain transformers, each query needs a one-sentence task instruction while documents need none. The 270m and 0.6b models also used distillation from larger embedders, and the 27b model has 5,376 dimensions.

Write a task instruction for queries
Encode queries and documents
Pool the last token and normalize
Compare vectors with a dot product
Rank passages by score

Title: Multilingual MTEB v2 score by model
harrier-oss-v1-270m: 66.5 points
harrier-oss-v1-0.6b: 69.0 points
harrier-oss-v1-27b: 74.3 points

Good to know:

• Context: all three sizes take up to 32,768 tokens, and the 27b model gives vectors of 5,376 dimensions.
• Size: about 54 GB in BF16, so it needs serious hardware.
• Languages: the record lists many, starting with af, am, ar, as, az, be, bg and bn.
• License: MIT.
• Cost to run: none named.
• Data it reaches: it downloads model files from Hugging Face, then works on the text you give it.
• Not for: writing text, since it outputs embeddings only.

On Hugging Face, 13 November 2026

• Task: feature extraction
• Library: sentence-transformers
• Parameters: 27B
• Weights: about 54 GB in BF16; about 13.5 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 7,115
• Likes: 173
• License: mit
• Languages: af, am, ar, as, az, be, bg, bn
• Quantized versions: 4
• Fine-tunes: 3
• Tags: sentence-transformers, safetensors, gemma3_text, feature-extraction, mteb, transformers, multilingual, text-embeddings-inference

Links

• Model: https://huggingface.co/microsoft/harrier-oss-v1-27b
• Files: https://huggingface.co/microsoft/harrier-oss-v1-27b/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download microsoft/harrier-oss-v1-27b

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