Microsoft's Bing team (yes, really) has released "Harrier," an open-source embedding model. Harrier supports more than 100 languages, offers a 32,000-token context window, and was trained on over two billion examples plus synthetic data from GPT-5. According to the team, Harrier takes the top spot on the multilingual MTEB v2 benchmark and outperforms proprietary models from OpenAI and Amazon.
| Rank (Borda) | Model | Zero-shot | Active Params (B) | Total Params (B) | Embedding Dim | Max Tokens |
|---|---|---|---|---|---|---|
| 1 | harrier-oss-v1-27b | 78% | 25.6 | 27.0 | 5376 | 131072 |
| 2 | KaLM-Embedding-Gemma3-12B-2511 | 73% | 10.8 | 11.8 | 3840 | 32768 |
| 3 | llama-embed-nemotron-8b | 99% | 7.0 | 7.5 | 4096 | 32768 |
| 4 | Qwen3-Embedding-8B | 99% | 6.9 | 7.6 | 4096 | 32768 |
| 5 | gemini-embedding-001 | 99% | 3072 | 2048 | ||
| 6 | Qwen3-Embedding-4B | 99% | 3.6 | 4.0 | 2560 | 32768 |
| 7 | Octen-Embedding-8B | 99% | 6.9 | 7.6 | 4096 | 32768 |
| 8 | F2LLM-v2-14B | 88% | 13.2 | 14.0 | 5120 | 40960 |
| 9 | F2LLM-v2-8B | 88% | 6.9 | 7.6 | 4096 | 40960 |
| 10 | harrier-oss-v1-0.6b | 78% | 0.440 | 0.596 | 1024 | 32768 |
Alongside the full 27-billion-parameter model, the team released two smaller variants—0.6B and 270M—designed to run on less powerful hardware. All three models are available on Hugging Face under the MIT license. Going forward, the team plans to integrate the technology into Bing and into new grounding services for AI agents.
Embedding models handle the searching, retrieving, and organizing of information AI systems need for accurate answers. According to Microsoft, they're becoming increasingly critical as AI agents independently take on more complex, multi-step tasks.
