Embed v1 4b

Embed v1 4b maps queries and documents into a shared vector space so retrieval reduces to approximate nearest neighbor search. Perplexity released it on February 26, 2026 as the larger of the two standard pplx-embed-v1 models. Embed v1 4b uses bidirectional attention with mean pooling over all token representations, rather than the causal attention that decoder-derived embedding models inherit.

On MTEB(Multilingual, v2), Embed v1 4b reaches an average nDCG@10 of 69.66% at INT8 precision, matching Qwen3-Embedding-4B at 69.60% and exceeding gemini-embedding-001 at 67.71%. On ToolRet, which measures retrieval over tool and API descriptions, it scores 44.45% average nDCG@10. On Perplexity's internal PPLXQuery2Query benchmark over a 2.4 million document corpus, it reaches 73.5% Recall@10 against 67.9% for Qwen3-Embedding-4B. On PPLXQuery2Doc over a 30 million page corpus, it reaches 91.7% Recall@1000 against 88.6%.

Two design choices shape how you integrate Embed v1 4b. It produces INT8 embeddings natively rather than as a post-hoc compression step, which cuts storage 4x compared with FP32; binary output cuts it 32x, and at this parameter scale Perplexity measures the binary quality drop at under 1.6 percentage points. Embed v1 4b also requires no instruction prefix, so you embed text directly. That removes a common failure mode where the instruction used at indexing time drifts from the one used at query time and quietly degrades recall. Matryoshka representation learning lets you request shorter vectors through the dimensions parameter when storage matters more than the last point of accuracy.