pplx-embed-v2-late-9b

Late interaction goes local. Perplexity’s pplx-embed-v2-late-9b (October 7, 2026, MIT) is an 8.39B late-interaction embedding model on the Qwen3.5 backbone: 32 layers, 4 KV heads x 256 head-dim, a 262,144-token context. Late-interaction retrieval keeps one vector per token instead of pooling into one vector, which buys recall on long documents at the cost of storing more vectors; the 9B size makes the trade affordable on a workstation rather than a cluster.

The MIT license does the differentiating. The v2 line’s smaller sibling (0.6B) covers the low end; against this site’s current local-RAG default, EmbeddingGemma 2 at 744M, the 9B late-interaction build is the upgrade path for retrieval stacks with money for an A100-class card or a 24GB unit at 16.8GB fp16. 467 downloads in the first days: early, but the license means no friction when the retrieval-benchmark posts arrive.