Advancing Agentic Information Retrieval with Decoder-based SoTA Embedding Models

We introduce a family of open-weight, commercially available multilingual text embedding models for production-scale retrieval. The models are adapted from decoder-only Ministral-3 backbones by converting causal self-attention into bidirectional self-attention and using mean pooling to produce fixed-size embeddings. To obtain an efficient 1B model, we compress a 3B embedding model through two rounds of retrieval-aware structured pruning and distillation from an 8B teacher. The models are trained with contrastive learning over mined query-passage pairs and fine-tuned on public and synthetic retrieval datasets. Across established public benchmarks, they achieve state-of-the-art retrieval accuracy while supporting RAG, agentic retrieval, code retrieval, memory, and long-context retrieval workloads.