Perplexity's pplx-embed-v2-late ships 0.6B and 9B multimodal embedding models: edge-ready size, hardware needs, and best 92.4% MADQA score.