Perplexity releases pplx-embed: embedding models that change the rules of search

Perplexity, which has transformed from a niche search startup into one of the most prominent players in the AI industry over the past two years, has taken another strategically important step. The company released pplx-embed — a family of multilingual embedding models that, according to developers, establish a new quality standard for information retrieval tasks across the entire internet. If Perplexity was primarily a consumer of other people's models before, it now increasingly asserts itself as a creator of its own infrastructure.

To understand the significance of this release, it's worth examining what embeddings are and why they are critically important. An embedding is a numerical representation of text in multidimensional space — a kind of mathematical fingerprint of meaning. When you enter a query into a search engine, it is the embedding model that determines which documents are semantically close to your question. The quality of this model directly affects how relevant the results you receive are. Until now, proprietary solutions from OpenAI, Cohere, and Google have remained the gold standard in this field, while among open models, developments from China and separate projects like Microsoft's E5 have led.

Pplx-embed is built on the Qwen3 architecture, but with a fundamental modification. Most modern language models use causal (unidirectional) attention — they read text left to right, as a person reads a book, and each token "sees" only what came before it. This works great for text generation, but for embedding tasks it is a serious limitation. When creating a holistic representation of a document, the model needs to account for context in both directions — both what comes before a word and what comes after it. Perplexity solved this problem by switching the architecture to bidirectional attention, essentially returning to ideas laid out in BERT, but at a qualitatively new level of scale and complexity.

The second key innovation is the use of a diffusion approach in the embedding creation process. Implementation details have not yet been fully disclosed, but the principle itself is borrowed from generative image models: instead of obtaining a text representation in a single pass, the model iteratively refines it, gradually "cleaning" it from noise. For working with real web data, which by definition is noisy — broken markup, ad insertions, duplicate content, a mix of languages — such an approach could be a decisive advantage. It is robustness to noise that distinguishes a model that performs well on clean benchmarks from a model that deals with real internet chaos.

The multilingual nature of pplx-embed deserves special attention. Qwen3, which underlies the model, was originally trained on data in over a hundred languages, and Perplexity apparently preserved and strengthened this property. For a company whose search product operates globally, this is not just a nice bonus but an operational necessity. A user from Tokyo, Moscow, or São Paulo should receive equally high-quality results, and a single multilingual embedding model is the most elegant way to achieve this.

The strategic context of this release is no less important than the technical aspects. Perplexity has long been dependent on external model suppliers — OpenAI for generation, various providers for embeddings. Each such dependency is both a financial risk and a ceiling for optimization. By releasing its own embedding models, Perplexity gains full control over a key link in its search pipeline. It can fine-tune models for its specific needs, optimize inference latency and cost, and most importantly, stop paying for each API call to competitors. For a company processing millions of search queries daily, savings could reach millions of dollars a year.

For the broader industry, this release signals an important trend: vertical integration in AI is accelerating. Companies that started as "wrappers" around other people's models are one by one beginning to build their own stacks. Perplexity follows a path already taken by others — from consuming APIs to creating their own models, from dependency to autonomy. The fact that the models are positioned as production-ready alternatives to proprietary APIs suggests that Perplexity is not just solving its internal tasks, but also eyeing the market for AI infrastructure services.

Pplx-embed is not a revolution but a logical evolutionary step, yet a most telling one. Perplexity demonstrates that it is ready to compete not only at the level of end-user product but also at the level of foundational technologies. If the claimed SOTA quality is confirmed by independent benchmarks, OpenAI and Google will face yet another serious competitor — and precisely where it hurts most: in the infrastructure upon which all modern AI search is built.