Microsoft's compact 0.6B multilingual embedder distilled for mid-tier and CPU deployments.
Access model weights, configuration files, and documentation.
See which devices can run this model and at what quality level.
The mid-tier Harrier model from Microsoft Research, a Qwen3-0.6B-based decoder-only embedding model trained with contrastive learning plus knowledge distillation from larger teacher embedding models. Supports 94 languages and 32K-token contexts and produces 1024-dim embeddings, all under the MIT license.
View on Hugging Face