pplx-embed-v2-late
Perplexity released pplx-embed-v2-late, two ColBERT-style multimodal embedding models at 0.6B and 9B, under the MIT license, with the 9B scoring 92.4% on MADQA.
The models keep a 128-dim vector for every token and score with MaxSim, instead of compressing a document into one vector. They retrieve text, images and rendered PDF pages in one shared space, so a 9B index can be queried with the 0.6B model. Perplexity distilled both from an 18B teacher.
- Date
- Wednesday 7 October 2026
- Lab
- Perplexity
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| MADQA accuracy (9B) | 92.4% 0.6B scores 90.1%; Mixedbread Agentic Search scores 93.4% | company |
| Domain-specific text nDCG@10, 72 tasks (9B) | 81.3% 0.6B scores 78.0% | company |
| BrowseComp+ accuracy (9B) | 64.0% 4.9pp above the next ColBERT model | company |
| ViDoRe v3 image nDCG@10 (9B) | 65.2% 0.6B scores 62.3%; Tencent EVIE scores higher | company |
All scores are self-reported by Perplexity and the technical report is not out yet. A hosted API endpoint is planned but not live.
Sources
This record was partly confirmed: some claims could not be checked on 8 October 2026. How we check