Harrier OSS v1
Microsoft's MIT-licensed multilingual text embedding family in 270M, 0.6B and 27B sizes, with 32K-token inputs.
In short. Harrier turns text in dozens of languages into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It comes in a tiny 270M version, a 0.6B one and a large 27B one, and Microsoft released all three for free under the MIT license.
Summary
Harrier OSS v1 is a family of multilingual text embedding models from Microsoft in three sizes: 270M, 0.6B and 27B parameters, producing vectors of 640, 1,024 and 5,376 dimensions from inputs of up to 32,768 tokens. The models are decoder-only, take the last token as the text's representation and normalize the vector, and Microsoft lists retrieval, clustering, semantic similarity, classification, bitext mining and reranking as uses. Training used contrastive learning on a large mixture of multilingual datasets, and the 270M and 0.6B models were also trained by distillation from larger embedding models. The card names 42 languages, including Arabic, Hindi, Japanese, Korean, Thai, Turkish, Vietnamese and Chinese, and says the support is not limited to them.
Microsoft reports Multilingual MTEB v2 scores of 66.5, 69.0 and 74.3 for the 270M, 0.6B and 27B models, and says they were state of the art on that benchmark on the release date. It publishes the evaluation prompts and points to the mteb repository for reproducing the scores. These are the developer's own numbers, and the card links no paper and does not describe the training data. Queries need a one-sentence task instruction in the form "Instruct: ... Query: ...", for which Sentence Transformers ships presets such as web_search_query, while documents are encoded without one. The BF16 files are 0.54 GB, 1.19 GB and 54.02 GB.
The license is MIT, and Microsoft's card names no base model. The configuration files show a Gemma 3 text architecture for the 270M and 27B models and Qwen3 for the 0.6B, and the MTEB leaderboard lists the 27B model as adapted from google/gemma-3-27b-pt, so check the upstream terms if your compliance team needs them. The models first appeared on Hugging Face on 30 March 2026. In the 30 days to 11 November 2026, Hugging Face counted about 222,000 downloads for the 0.6B model, 105,000 for the 270M and 7,400 for the 27B.
Good for
- Multilingual semantic search and retrieval-augmented generation over long documents, with inputs up to 32,768 tokens and an MIT license
- A small 270M or 0.6B embedder that runs on modest hardware, with the 27B model for the highest reported accuracy
- Clustering, classification and bitext mining as well as retrieval, all from one model family
Watch out for
- Queries must carry a one-sentence task instruction and documents must not, because the card says that is how the models were trained
- The scores are Microsoft's own, and the card gives no paper, training-data description or base model
- The 27B model has 5,376-dimension vectors and a 54 GB file, which makes storage, indexing and serving much heavier than the smaller two
- It handles text only, and the card lists 42 languages without per-language scores
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Harrier OSS v1. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| 270M (640-dimension vectors) BF16 | 270M | 0.54 GB | 0.621 GB | Laptop or phone |
| 0.6B (1,024-dimension vectors) BF16 | 600M | 1.19 GB | 1.37 GB | Laptop or phone |
| 27B (5,376-dimension vectors) BF16 | 27B | 54 GB | 62 GB | Workstation |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice.
See also
Frequently asked questions
- What is Harrier OSS v1?
- Harrier turns text in dozens of languages into number vectors that a search system can compare, so it can find passages that mean the same thing as a question. It comes in a tiny 270M version, a 0.6B one and a large 27B one, and Microsoft released all three for free under the MIT license.
- Can I use Harrier OSS v1 commercially?
- Yes. Harrier OSS v1 is released under MIT, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Harrier OSS v1 open source?
- Harrier OSS v1 is released under MIT, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run Harrier OSS v1 on my own computer?
- Yes. The smallest variant, 270M (640-dimension vectors), needs about 0.621 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Harrier OSS v1 released?
- Harrier OSS v1 was released by Microsoft on March 30, 2026.
- How long a context window does Harrier OSS v1 have?
- Harrier OSS v1 supports a context window of up to 32K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- harrier-oss-v1-0.6b model card (Hugging Face) huggingface.co
- harrier-oss-v1-27b model card (Hugging Face) huggingface.co
- harrier-oss-v1-270m model card (Hugging Face) huggingface.co
- harrier-oss-v1-27b on the MTEB leaderboard leaderboard.mteb.org
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Qwen3-Embedding Alibaba Qwen | 600M – 7.57B | 32K | Apache 2.0 (Permissive) | 1.38 GB Laptop or phone | |
| BGE-M3 BAAI | 568M | 8K | MIT (Permissive) | 2.61 GB Laptop or phone | |
| EmbeddingGemma 2 Google DeepMind | 270M – 740M | 8K | Apache 2.0 (Permissive) | 0.621 GB Laptop or phone | |
| Jina Embeddings v5 Jina AI | 239M – 677M | 32K | CC BY-NC (Non-commercial) | 0.552 GB Laptop or phone |