The Pulse
Perplexity Releases 0.6B and 9B Multimodal Embedding Models
Perplexity released two publicly available embedding models that search text, images and rendered PDF pages using token-level vectors. The 0.6B and 9B models share an embedding space, allowing a smaller model to query an index built by the
AI.info Team · Written with AI ·
Perplexity’s new 0.6-billion-parameter embedding model can query an index built by its 9-billion-parameter counterpart, letting developers reserve the heavier computation for document indexing and use the smaller model for live searches. The company released the pair, named pplx-embed-v2-late, on October 6, with support for text and images, including rendered PDF pages. Both models are publicly available on Hugging Face; Perplexity says support on its API platform will arrive progressively.
Two model sizes, one retrieval space
The models use late interaction: rather than compressing a document or query into one vector, they retain a vector for each input token. Perplexity says the ColBERT-style models produce 128-dimensional token embeddings and score a query against a document with MaxSim. For each query token, the system finds the highest-scoring document token, then adds those similarities to produce a relevance score.
“However, rather than reducing each input to a single vector, they retain multiple vectors and use a richer scoring function when comparing a query with a document.”
Perplexity and the listed authors of “Multimodal embeddings beyond a single vector”
Keeping more detail can help a search match separate parts of a query to different parts of a document. It also carries a cost: the company says late-interaction retrieval stores vectors for document tokens and compares many token pairs, so storage and candidate-scoring costs rise with document length. That makes it a different operating trade-off from dense retrieval, which stores one vector per document and uses a single inner product to score a candidate.
Searching a page without first extracting its text
The models support text-to-text and text-to-image retrieval, including searches over rendered PDF pages without first running OCR or parsing the page into text. Perplexity argues that conventional extraction can lose information such as layout, table structure, figures and spatial relationships, or introduce OCR errors. The same models also support semantic search over natural images.
The smaller model is designed for latency-sensitive use and can run on edge devices, according to Perplexity. It has 594 million parameters in total, but the company says only about 240 million are active for text encoding and 340 million for image encoding; its parameter count also includes a 254-million-parameter token-embedding table that Perplexity says has little effect on runtime. The 9B model is intended for higher-quality document indexing, while the shared embedding space lets the 0.6B model handle queries against that index.
Perplexity’s benchmark claims—and their limits
Perplexity reports that its 0.6B model matched models with five times as many active parameters on the ViDoRe V3 visual-document retrieval benchmark. On the company’s 72 domain-specific text tasks, it averaged 78.0%, compared with 81.3% for the 9B model. Those are Perplexity’s own evaluations; the company says it excluded datasets associated with the benchmarks from its training data.
For a large-scale web-retrieval test called Q2D-Web, Perplexity reports scores of 74.8% for the 9B model and 73.6% for the 0.6B model, measured as Recall@1000 on the combined relevance judgments. The company says the benchmark uses about 70,000 agent-reformulated queries drawn from production search traffic and a corpus of 190 million web documents. It also notes that its combined judgments include language-model assessments of documents its own production systems did not retrieve, addressing a potential bias in judgments based only on those systems’ results.
Weights are available; API access is still to come
The Hugging Face releases work with sentence-transformers version 6.0.0 or later and transformers version 5.4.0 or later, according to Perplexity. The company says both checkpoints are available now, but its post describes API support as a future, progressive rollout rather than a feature already available. Perplexity also says it plans to publish a technical report with further details and complete evaluation results later this year.