What is Harrier-OSS-v1 (270M, 0.6B, 27B)?
Harrier-OSS-v1 is a family of open multilingual text embedding models from Microsoft, released on 30 March 2026 under the MIT licence in three sizes: harrier-oss-v1-270m, harrier-oss-v1-0.6b and harrier-oss-v1-27b. An embedding model turns text into a list of numbers so that similar meanings land close together, for search, clustering and classification. Microsoft's Bing team says Harrier builds on its earlier E5, Multilingual E5 and GritLM work.
All three are decoder-only models that take the last token's state as the embedding, read up to 32,768 tokens and cover more than 100 languages. Vector sizes are 640, 1,024 and 5,376 dimensions. The 270M and 27B use Google's Gemma 3 architecture and the 0.6B uses Qwen3; the two smaller models were distilled from larger ones. Training used over 2 billion weakly supervised pairs and over 10 million curated ones, with synthetic data from GPT-5.
Is Harrier-OSS-v1 (270M, 0.6B, 27B) worth it?
Harrier-OSS-v1 tops Microsoft's multilingual text embedding comparisons at every size it ships. Microsoft reports 74.3 on multilingual MTEB v2, an average over 131 tasks, for the 27B, first on the leaderboard as of 6 April 2026 and 2 points above the previous best of 72.3. In the same comparison OpenAI's text-embedding-3-large scored 58.92 and Google's multimodal Gemini Embedding 2 69.9, so even the 0.6B, at 69.0, beats most of the paid APIs in Microsoft's table.
The small models matter more for most buyers. The 0.6B scores 69.0 against 64.3 for Qwen3-Embedding-0.6B, and the 270M 66.5 against 61.2 for EmbeddingGemma-270m. IBM's own comparison in its Granite R2 paper agrees: on multilingual MTEB retrieval the 270M scored 66.4, ahead of every other model under 500 million parameters it tested, including IBM's. The 0.6B logged about 230,000 Hugging Face downloads in the past month, the 27B about 7,450.
The 27B is a research-grade flagship. At 27 billion parameters it needs about 54 GB just for its weights in BF16, its 5,376-dimension vectors are five times the storage of the 0.6B's 1,024, and it is text only. Most teams will treat it as a teacher or an offline reranking aid rather than the model behind every query.
The catch: Microsoft offers no hosted endpoint, so there is no pay-per-token option and you run it yourself, on Azure through Foundry or elsewhere. Queries need a one-sentence task instruction or accuracy drops, the card mentions no way to shorten vectors, and the leaderboard figures are Microsoft's runs with the prompts it published.
Our take: Harrier 0.6B is the multilingual text embedder to benchmark first on your own data in 2026; reach for the 27B only if a few points of recall are worth a large GPU.
Where it earns its place
A search engineer at a European marketplace with listings in 20 languages
Let shoppers search in their own language and find listings written in another
Embed listings with harrier-oss-v1-0.6b on a single GPU and encode queries with the web_search_query prompt; 1,024-dimension vectors keep the index modest, and MIT avoids any licence review.
An ML team at a bank building an internal knowledge assistant
Improve retrieval quality without sending documents to an embedding API
Self-host the 0.6B or 270M on existing hardware, or deploy from the Foundry catalogue inside the bank's Azure tenant; documents never leave the bank's own compute.
A research group building a domain-specific embedder
Create training labels for a smaller model in their field
Use the 27B offline as a teacher, the role Microsoft gave it for the smaller Harrier models, to score query-document pairs, then fine-tune a compact model for production.
Before you commit
- Setup
- pip install sentence-transformers, load microsoft/harrier-oss-v1-0.6b and encode queries with prompt_name='web_search_query' (or a custom 'Instruct: ... Query:' prompt) and documents with no prompt.
- Integrations
- Sentence Transformers and Transformers; the mteb package with published evaluation prompts; one-click deployment from Hugging Face into Microsoft Foundry.
- Data and privacy
- MIT licence, so weights can be self-hosted and modified freely; documents stay on your hardware. Microsoft says training data included synthetic examples generated by GPT-5.
- Cost at scale
- Weights are free. The 270M and 0.6B run on modest GPUs; the 27B needs about 54 GB of accelerator memory in BF16 before activations, and its vectors take 5.25 times the storage of the 0.6B's.
- Limits
- Text only; 32,768 tokens per input; last-token pooling means padding side matters if you skip Sentence Transformers; queries need an instruction; fixed vector sizes of 640, 1,024 or 5,376.
What’s good
- 27B first on multilingual MTEB v2 at 74.3 (6 April 2026)
- 0.6B at 69.0 and 270M at 66.5 lead their size classes
- MIT licence for all three sizes
- 32,768-token input, more than 100 languages
What’s not
- Text only
- No hosted API from Microsoft
- 27B needs about 54 GB for weights and stores 5,376-dimension vectors
- Queries need an instruction prompt
- No documented vector truncation
Best for
Teams that self-host multilingual text search and want the best open accuracy per parameter with no licence strings.
Harrier-OSS-v1 (270M, 0.6B, 27B) alternatives
Harrier-OSS-v1 (270M, 0.6B, 27B): common questions
- Is Harrier-OSS-v1 (270M, 0.6B, 27B) worth it?
- Harrier is the most accurate open multilingual text embedder on the leaderboard that matters, and its small sizes are the better story. We rate it 4.5 out of 5.
- How much does Harrier-OSS-v1 (270M, 0.6B, 27B) cost?
- Free under MIT. No hosted API from Microsoft; the models can be deployed on your own Azure compute from the Foundry catalogue, billed as normal compute.
- What are the downsides of Harrier-OSS-v1 (270M, 0.6B, 27B)?
- Text only; no hosted API from Microsoft; 27B needs about 54 GB for weights and stores 5,376-dimension vectors; queries need an instruction prompt; no documented vector truncation.
- Who is Harrier-OSS-v1 (270M, 0.6B, 27B) best for?
- Teams that self-host multilingual text search and want the best open accuracy per parameter with no licence strings.
- How much does Harrier-OSS-v1 (270M, 0.6B, 27B) cost for a team or in production?
- Weights are free. The 270M and 0.6B run on modest GPUs; the 27B needs about 54 GB of accelerator memory in BF16 before activations, and its vectors take 5.25 times the storage of the 0.6B's.
- Is Harrier-OSS-v1 (270M, 0.6B, 27B) safe for company data?
- MIT licence, so weights can be self-hosted and modified freely; documents stay on your hardware. Microsoft says training data included synthetic examples generated by GPT-5.
- What are the best Harrier-OSS-v1 (270M, 0.6B, 27B) alternatives?
- The closest alternatives we have reviewed are Qwen3-Embedding and Qwen3-Reranker, Nemotron 3.5 Embed (1B and 8B), Granite Embedding Multilingual R2 (311M and 97M), EmbeddingGemma 2 and Jina Embeddings v5.
Answers checked Nov 7, 2026.
Make Harrier-OSS-v1 (270M, 0.6B, 27B)? Show this rating on your site
