Two sizes, one space
Perplexity released pplx-embed-v2-late on 7 October: two open-weight retrieval models, at 0.6B and 9B parameters, published on Hugging Face under an MIT licence. They are late-interaction retrievers in the ColBERT mould, emitting a 128-dimensional vector for every token rather than compressing a document into one vector, and scoring a query against a document with MaxSim — each query token picks its best-matching document token, and those maxima are summed.
The feature Perplexity leads with is not a benchmark number but an interoperability property. Both checkpoints were distilled from the same 18B teacher using what the company calls LEAF-style representation distillation, aligning each student’s per-token output with the teacher’s. Because of that, the two models share an embedding space, and a corpus indexed by the 9B model can be searched with queries embedded by the 0.6B one.
Why that is the point
Indexing and querying have different economics. Indexing runs offline, once, and its cost is spread over every subsequent search; query encoding runs on the latency path of every request. Perplexity’s argument is that you should pay for quality where it is cheap and for speed where it is not.
The company reports what that asymmetric setup buys. On ViDoRe v3 image retrieval, indexing with the 9B model and querying with the 0.6B scores 63.5%, up from 62.3% when the small model does both — a gain of 1.2 percentage points at identical query cost. Across 72 domain-specific text benchmarks the asymmetric configuration averages 1.6 points better. Running the 9B model on both sides is better still, at 81.3% on those domain benchmarks against the 0.6B’s 78.0%, so the mixed setup recovers roughly half the gap for free at query time.
The numbers, and whose they are
Every figure below is Perplexity’s own, published alongside the release; none is independently verified.
On the 72 domain-specific benchmarks — finance, legal, health, conversation, tech, plus multilingual tasks — the 9B model averages 81.3%, which the company says is the best of everything it evaluated and 1.6 points ahead of the next model. The 0.6B averages 78.0%, 0.3 points behind gemini-embedding-2. On ViDoRe v3 Markdown the two score 64.7% and 61.2%, against 60.6% for the best external model tested.
On Q2D-Web, Perplexity’s own benchmark of about 70,000 agent-reformulated queries over 190 million web documents, both models lead every baseline on all three relevance judgement sets; on the combined set they reach 74.8% and 73.6% recall@1000 against a previous best of 69.3%. On visual retrieval, the 9B scores 65.2% on ViDoRe v3 image, trailing only Tencent’s EVIE among the models compared; the 0.6B reaches 62.3% while activating 340M parameters for image input.
On agentic tasks, a GPT-OSS-120B agent using the 9B retriever scores 64.0% on BrowseComp+, which Perplexity says is 4.9 points ahead of the next ColBERT model and 8.7 ahead of the next dense one. On MADQA — 500 human-written questions over 800 real PDFs spanning more than 18,000 pages — a Gemini 3.5 Flash agent paired with the 9B model reaches 92.4% accuracy, against 88.9% with Mixedbread’s retriever. Perplexity notes that Mixedbread’s full agentic search system scores 93.4%, within its confidence interval.
The costs the post does not hide
Late interaction stores vectors per token, not per document, so the index is larger and candidate scoring is heavier, and the cost grows with document length. Perplexity says so plainly, and frames the 128-dimensional output as a storage advantage against competitors at 2,048 and 4,096 dimensions.
The company also says it excluded from training any dataset associated with a benchmark it evaluates on, accepting a quality cost to keep the numbers meaningful. A hosted API is not yet available; the models are for self-hosters today, with Perplexity saying it will roll out late-interaction, dense and contextual embeddings on its platform over time, and publish a technical report later this year.
