Harrier: Open embeddings
Microsoft’s Bing team open‑sourced Harrier, a multilingual embedding model that supports over 100 languages and long 32K inputs, positioning it for semantic search and retrieval‑augmented apps. Harrier already ranks at the top of the multilingual MTEB‑v2 leaderboard, making it a practical choice for builders who need high‑quality embeddings without vendor lock‑in. (x.com)
Search engines do not read text the way people do. They turn each passage into a list of numbers, like a map coordinate, so “cheap flights to Tokyo” lands near pages about airfare even if the exact words do not match. (blogs.bing.com) That numeric map is called an embedding model, and it is the quiet layer underneath semantic search, recommendation systems, and the document lookup step inside chatbots. Microsoft’s Bing team just released its own version, called Harrier, as open source instead of keeping it behind an application programming interface meter. (blogs.bing.com) Harrier comes in three sizes, with 270 million, 600 million, and 27 billion parameters, which are the adjustable weights the model uses to place text in that map. Microsoft published all three on Hugging Face under an Massachusetts Institute of Technology license, which allows commercial use with few restrictions. (huggingface.co) The practical hook is language coverage. Microsoft says Harrier works across more than 100 languages, and outside reporting on the model cards describes 94 languages in the released family, which puts it in the range needed for a global search product instead of an English-only demo. (blogs.bing.com, aihola.com) The other unusual number is 32,768 tokens of input length. A token is a chunk of text smaller than a word, and a 32,768-token window means the model can embed long manuals, legal filings, or multi-page support documents without chopping them into as many pieces first. (huggingface.co) Microsoft says Harrier ranks first on the multilingual version 2 Massive Text Embedding Benchmark leaderboard as of April 6, 2026. That benchmark is the industry’s broad report card for embeddings, testing retrieval, clustering, classification, and similarity tasks across many datasets instead of one cherry-picked example. (blogs.bing.com, huggingface.co) On the flagship model card, Microsoft lists a multilingual Massive Text Embedding Benchmark version 2 score of 74.3 for the 27 billion parameter model, 69.0 for the 600 million model, and 66.5 for the 270 million model. That spread gives developers a tradeoff between accuracy and hardware cost instead of one giant model that only fits in a lab. (huggingface.co) Microsoft trained Harrier with contrastive learning, which means it learns by pulling related texts closer together and pushing unrelated texts farther apart in vector space. The smaller two models were also trained by distillation, where a larger model acts like a teacher and compresses some of its behavior into a cheaper student. (huggingface.co) This release lands in a crowded market where builders often rent embeddings from OpenAI, Google, Cohere, or cloud search stacks and pay every time they process text. Open weights change that math because a company can run the model on its own infrastructure, keep sensitive documents inside its own network, and avoid being locked to one vendor’s pricing. (blogs.bing.com, huggingface.co) It also says something about where Microsoft thinks the next fight is. Large language models write the answer, but embedding models decide which facts the system sees first, and Bing is now giving away the retrieval layer it says helped it reach the top of the benchmark table. (blogs.bing.com, x.com)
