Perplexity open-sources multimodal multi-vector embedding model
What happened
On October 8, 2026, Perplexity open-sourced pplx-embed-v2-late in 9B and 0.6B versions, with weights on Hugging Face. Both are late-interaction multi-vector embedding models that handle text and images and share an embedding space. The 9B version can index multimodal data; the 0.6B version runs queries on-device and retrieves PDF pages without OCR. The release materials report 92.4% on MADQA and 64% on BrowseComp+.
Written by AI from the coverage · updated 2 hours ago
Coverage
Follow the reports to see the story from different sides.
- AI HOT · ModelsPerplexity 开源 pplx-embed-v2-late 多模态嵌入模型，提供 9B 与 0.6B 版本
Perplexity 开源 pplx-embed-v2-late，提供 9B 和 0.6B 两个文本与图像 late-interaction 多向量嵌入模型，权重已在 Hugging Face 提供。两者共享嵌入空间，9B 可用于多模态数据索引，0.6B 可在设备端查询，无需 OCR 即可检索 PDF 页面。材料报告模型在 MADQA 得分 92.4%，在 BrowseComp+ 得分 64%。
Heat over time
Not enough continuous observations to draw a trend yet.