Perplexity has released pplx-embed-v2-late, a new pair of ColBERT-style multimodal embedding models available in 0.6B and 9B sizes under the MIT license. Both models are designed to retrieve text, images, and rendered PDF pages within a single shared embedding space, eliminating the need for separate optical character recognition steps. The smaller 0.6B model is built for fast queries and edge device deployment, while the 9B model focuses on maximum quality, achieving a score of 92.4 percent on MADQA in self-reported evaluations.
The architecture uses token-level vectors measuring 128 dimensions, relying on MaxSim scoring to match query tokens to document tokens. Users can cross-mix model sizes by searching a 9B document index using the lightweight 0.6B query encoder to recover quality while keeping costs low. While the models demonstrate strong capabilities in visual document search and agentic retrieval-augmented generation, limitations include index sizes that grow with document length and a lack of support for mixing text and images in a single input.
Perplexity published the model weights on Hugging Face for commercial use, though a hosted API endpoint and technical report are still forthcoming. These models provide developers and enterprises with open-source tools for high-performance multimodal document retrieval and low-latency search applications.