Perplexity Open-Sources pplx-embed-v2-context Embedding Models, 9B Sets Retrieval SOTA
On November 3, Perplexity open-sourced its contextual embedding model pplx-embed-v2-context on Hugging Face, offering 9B and 0.6B open-weight versions along with the paper "Contextual Embeddings Beyond the Gold Chunk." The series is fine-tuned from pplx-embed-v2-late-9b; according to Perplexity's antoinechaffin, the 9B model scores 45.5 Answer@10 on turbopuffer's context-bench, significantly surpassing voyage and setting a new retrieval SOTA.
Confirmed
- Open-sourced on Hugging Face in 9B and 0.6B versions, fine-tuned from pplx-embed-v2-late-9b, built on the Qwen3.5 architecture (per posts from @AravSrinivas and antoinechaffin)
- The 9B model scores 45.5 Answer@10 on context-bench, beating voyage
- Training pipeline: first train a 9B teacher with expanded data, then produce the 0.6B model via the same representation-level distillation used for the late-interaction model; both share one embedding space, enabling cross-querying
- Training objective: combines a chunk-level KL loss with a document-level contrastive loss using max-pooled chunk scores, so the model learns both which chunks contain answers or supporting evidence and how to locate the right document among distractors
Why it matters
Context-aware embeddings take retrieval beyond isolated chunk matching to understanding the relationship between answers and context; the shared embedding space of the 9B/0.6B pair lets you mix large and small models in the same index, balancing quality and cost — directly useful for RAG and search.
2026-11-03 ~ 2026-11-03 · 5 related posts
Primary sources
- Perplexity's 9B contextual embedding sets SOTA, scores 45.5 on context-bench vs 31.1 for voyage — antoine_chaffin ·
- Perplexity open-sources pplx-embed-v2 contextual embedding models in 9B and 0.6B sizes — antoine_chaffin ·
- Perplexity distills 9B embedding teacher into 0.6B model sharing the same embedding space — antoine_chaffin ·
- Perplexity releases open-weight contextual embedding models (9B/0.6B), sets SOTA on ConTEB — AravSrinivas · 2026-11-03
- Training combines chunk-level KL loss with document-level contrastive loss on pooled scores — antoine_chaffin · 2026-11-03
- [source] Perplexity distills 9B embedding teacher into 0.6B model sharing the same embedding space — antoine_chaffin · 2026-11-03
- [source] Perplexity's 9B contextual embedding sets SOTA, scores 45.5 on context-bench vs 31.1 for voyage — antoine_chaffin · 2026-11-03
- [source] Perplexity open-sources pplx-embed-v2 contextual embedding models in 9B and 0.6B sizes — antoine_chaffin · 2026-11-03