Microsoft · 2026-04-06 · major
Harrier — Microsoft's Open-Source Multilingual Embedding Model Family
MIT-licensed family of multilingual embedding models (270M, 0.6B, 27B) from Microsoft's Bing team, ranking first on multilingual MTEB-v2 with 32k context and 100+ language support.
Microsoft's Bing team open-sources a decoder-only embedding family that tops the multilingual MTEB-v2 leaderboard.
Key specs
| License | MIT |
|---|---|
| Mteb v2 (27 b) | 74.3 |
| Sizes | 270M / 0.6B / 27B |
| Languages | 100+ |
What is it?
Harrier-OSS-v1 is a family of three MIT-licensed text embedding models from Microsoft's Bing team. Available in 270M, 600M, and 27B parameter sizes, each model takes text in any of 100+ languages and produces dense vector representations for retrieval, clustering, classification, and semantic similarity. The 27B variant ranks first on the multilingual MTEB-v2 benchmark.
How does it work?
Unlike most embedding models that use encoder-only architectures, Harrier is decoder-only with last-token pooling and L2 normalization -- the same transformer family as generative LLMs. This lets it leverage scaling laws that made GPT-class models powerful. All variants share a 32k-token context window and were trained on over 2 billion examples plus synthetic data.
Why does it matter?
Embeddings are the foundation of RAG pipelines, semantic search, and agent memory. Having a state-of-the-art multilingual option under MIT license means teams can run retrieval locally without API costs or vendor lock-in. The decoder-only architecture also signals a convergence between embedding and generation model families.
Who is it for?
ML engineers building RAG pipelines, search systems, or agent memory layers.
Try it
pip install sentence-transformers && model = SentenceTransformer('microsoft/harrier-oss-v1-0.6b')