Perplexity unveils new embedding models with enhanced performance on benchmark tasks

Perplexity AI has launched two new ColBERT-style multimodal embedding models, pplx-embed-v2-late, designed for efficient text, image, and PDF retrieval. The models come in two versions: a 0.6B parameter edge model for fast and cost-effective queries, and a 9B parameter model offering higher quality. Both models operate within a shared embedding space and are available on Hugging Face under the MIT license. The 9B model achieved a 92.4% score on the MADQA benchmark, outperforming other models by a significant margin. The 0.6B model, while scoring lower on certain benchmarks, still ranks second on ViDoRe v3 Markdown. Perplexity claims the models were trained using an 18B teacher model with LEAF-style token-level training, enabling their shared embedding space. The company also announced plans for a hosted API endpoint, though it is not yet live.