Perplexity open-sources pplx-embed-v2: skip OCR, index PDFs and images directly
lmoroney · x · 2026-10-08
Perplexity released pplx-embed-v2-late on Hugging Face under MIT license, in 0.6B and 9B sizes built on Qwen3.5, distilled from an internal 18B teacher.
- ColBERT-style late-interaction: every token gets a 128-dim vector, scored via MaxSim per-token matching
- Handles text, images, and rendered pages (PDFs, slides) directly — no OCR step needed
- Both sizes share one embedding space, so the 0.6B can query an index built with the 9B
- ViDoRe v3 image track: nDCG@10 of 62.3% (0.6B) and 65.2% (9B)
Caveat: per-token vectors make indexes large — test on a few hundred of your own documents first.
More from Models
- Anthropic launches Project Glasswing: Claude Mythos Preview hunts software bugs — zetalyrae · 2026-10-08
- OpenAI Paper Joked to Solve 700 Open Math Problems, Zero Verified Operating Systems — avaitopiper · 2026-10-08
- Skeptical of 'small model for evals' moats: labs will distill it into cheaper models — Shahules786 · 2026-10-08
- An Uber driver's take on OpenAI usage limits: they're bonuses, and bonuses can be cut — Arqium · 2026-10-08
- Bengio disputes 'just a sandbox bug' framing of AI agent hacks in FT op-ed — AlexTensor · 2026-10-08
- OpenAI model proves Hilbert's Tenth Problem false over Q, sidestepping 80-year approach — aran_nayebi · 2026-10-08