Perplexity has released two multimodal embedding models that refuse to reduce an entire document to one vector. The 0.6-billion and 9-billion-parameter models retain a 128-dimensional embedding for each token and use a late-interaction scoring method. Text queries can search text, images and rendered document pages in a shared space, including indexes produced by the larger model and queried by the smaller one.[1][2][3]
The design tackles an information bottleneck. A long report may discuss many entities, periods and exceptions. Compressing all of that into one fixed-size vector asks the representation to preserve every future reason someone might search for the document. Token-level vectors allow one query term to match one passage and another term to match a different part of the page.
What the late-interaction family changes
| Measure | Value | Context |
|---|---|---|
| Released sizes | 0.6B and 9B | The models target different quality, latency and compute requirements.Source 1 |
| Representation | 128 dimensions per retained token | The system stores many vectors for a document rather than one fixed-size vector.Source 1Source 2 |
| Modalities | Text and images | Rendered PDF pages can be searched directly without relying only on OCR text.Source 1 |
| Shared space | Small queries can search a large-model index | Index-time compute can be separated from latency-sensitive query encoding.Source 1 |
| Active text parameters in small model | About 240 million | Perplexity says nominal parameter count overstates text serving cost.Source 1Source 2 |
The storage trade-off is large enough to shape the product
Assume 32-bit floating-point storage. One token vector contains 128 values, requiring 512 bytes. A document retaining 1,000 token vectors therefore needs about 512,000 bytes, or 0.512 MB, before metadata, compression and index structures. One million documents of that size would require roughly 512 GB of raw vectors.
For comparison, one dense 768-dimensional float32 vector uses 3,072 bytes per document, or about 3.072 GB for one million documents. Under those simplified assumptions, the 1,000-token late-interaction representation uses around 167 times as much raw vector storage. Real systems can quantize, prune or compress vectors, and documents vary in length; the arithmetic explains why richer retrieval needs a different serving design.
Single-vector and late-interaction retrieval
| Feature | Single dense vector | Token-level late interaction |
|---|---|---|
| Document representation | One fixed-size point | Many vectors preserving local detail.Source 1 |
| Candidate scoring | One inexpensive similarity operation | Many token-to-token comparisons combined with MaxSim.Source 1 |
| Index size | Relatively compact | Grows with retained document length.Source 1 |
| Multimodal page search | Can encode rendered pages if the chosen model supports images; a single vector compresses the page | Can preserve separate token-level matches within an encoded pageSource 1 |
| Typical role | Cheap first-stage retrieval at very large scale | Quality-sensitive first stage or a richer reranking stage.Source 1 |
Shared embeddings create asymmetric deployment options
Indexing is usually an upfront cost, repeated as documents change or an incompatible embedding model replaces the old one. Query encoding happens on every search and is latency-sensitive. Perplexity’s shared space allows documents to be embedded by the 9B model while live queries use the 0.6B model. That can spend more compute on the durable corpus while keeping the interactive path lighter. It also permits local–cloud combinations, provided the data path is specified.[1]
Compatibility does not remove privacy questions. A local query representation may still reveal information when sent to a cloud index. A local document index may be copied into backups. Product teams need to describe where raw documents, token vectors, queries, candidate results and final answers are processed and retained.
Retrieval quality is not answer quality
A better retriever can put the right page in front of an agent more often. The agent can still misread the table, combine the wrong periods or ignore a decisive exception. Evaluation should therefore separate retrieval—whether relevant evidence appears near the top—from answer grounding—whether the system uses that evidence correctly. The distinction becomes especially important when visual documents contain charts, footnotes and spatial relationships.
Perplexity’s release is significant because it makes the compression trade-off explicit. Search can preserve more of a document’s structure, but the extra detail has an infrastructure cost. The best deployment will not simply select the highest benchmark. It will choose the representation that retrieves the decisive evidence often enough for the storage, latency and privacy budget available.
Sources & further reading
- Multimodal embeddings beyond a single vector Perplexity · 7 Oct 2026 · Accessed 8 Oct 2026
- pplx-embed-v2-late-0.6b model card Perplexity on Hugging Face · 7 Oct 2026 · Accessed 8 Oct 2026
- pplx-embed-v2-late-9b model card Perplexity on Hugging Face · 7 Oct 2026 · Accessed 8 Oct 2026
- Perplexity Research Perplexity · 8 Oct 2026 · Accessed 8 Oct 2026