MicrosoftのBingチームが「Harrier」埋め込みモデルをオープンソース化

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.

AI News Without the Hype – Curated by Humans

Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.

Subscribe now