Perplexity Releases pplx-embed-v2-late: Multi-Vector Embeddings for Text and Visual Documents

Perplexity has introduced pplx-embed-v2-late, a collection of late-interaction embedding models available in 0.6B and 9B parameter sizes. These models utilize a ColBERT-style architecture with 128-dimensional token embeddings and a MaxSim scoring function to enable fine-grained matching between query and document parts. The models support both text and image modalities, allowing for direct retrieval from visual documents without traditional OCR pipelines. Training involved distillation from an 18B teacher model using LEAF-style representation distillation on 186 million query-document pairs across 594 datasets. The 0.6B model is based on Qwen3.5-0.8B, with its text tower pruned to 12 layers. Both models share an embedding space, enabling asymmetric retrieval where a corpus indexed by the 9B model can be queried by the 0.6B model. The models are available on Hugging Face and compatible with sentence-transformers >= 6.0.0 and transformers >= 5.4.0.