According to CoinMeta, Perplexity has open-sourced two multimodal embedding models, with capacities of 0.6B and 9B respectively. The new models support text, image, and PDF page retrieval, and the vectors of both models are compatible with each other. Developers can use the 9B model to build a database and then employ the 0.6B model for routine searches. Each token is assigned a 128-dimensional vector in the new models, which allows for better matching of relevant content within documents. In official Vidore V3 image retrieval tests, it was found that using the 0.6B model for database construction and queries resulted in a score of 62.3%; switching to the 9B model for database construction and using the 0.6B model for queries increased the score to 63.5%. Using the 9B model exclusively yielded a score of 65.2%. The weights of both models have been made available on Hugging Face according to the MIT protocol.