Announced with turbopuffer in Contextual embedding beyond the gold passage, the student starts from a 9B ColBERT retriever, projects to 2048 dimensions, supports Matryoshka 1024-d and native int8, and is trained with a document-level contrastive loss plus distillation from a query-aware context compressor instead of a single gold-chunk label. A Hugging Face preview is public; Perplexity said the API was still forthcoming. Reported ConTEB and private context-bench scores are vendor-run.