Microsoft’s Bing team has open sourced Harrier under the MIT licence, making the first open embedding family to top multilingual MTEB v2 available for unrestricted enterprise use.
Microsoft’s Bing team has open-sourced Harrier, a family of embedding models released under the MIT licence with no usage restrictions, marking the first open-source embedding stack to top the multilingual MTEB v2 benchmark. Available on Hugging Face for commercial use, modification, and redistribution, the family supports more than 100 languages and directly outperforms proprietary alternatives from OpenAI, Google, Amazon, and previous NVIDIA leaders.
The release strengthens the view that proprietary embedding models no longer retain a clear quality edge, accelerating the commoditisation of foundation AI infrastructure for enterprise retrieval, RAG, legal discovery, and scientific search.
Harrier ships in three tiers: the flagship 27B model with 25.6B active parameters and 5,376-dimensional embeddings, a 0.6B mid-range version, and a 270M edge model. All three share the same API, embedding format, and a 32,768-token context window, enabling a consistent path from prototyping to production-scale deployment.
Microsoft also plans to integrate Harrier across Bing, grounding services, and AI agent memory, ranking, and orchestration layers. The Bing team described grounding as “the foundational capability that drives user trust for any AI agent”, positioning Harrier as “a foundational layer for memory, ranking, and orchestration” in the emerging agentic web.
For enterprises, the MIT-licensed release offers a direct open alternative to vendor-controlled embedding APIs while reducing lock-in, pricing volatility, and terms-of-service risks.
