Perplexity AI Releases pplx-embed-v2-late: A 0.6B Edge Model and a 9B Model Scoring 92.4% on MADQA

- Perplexity dropped pplx-embed-v2-late, splitting the difference between a 0.6B edge model and a 9B indexer. It’s MIT-licensed, so you can self-host without begging Core Dynamics for permission. The 9B model hits 92.4% on MADQA, which is impressive until you realize it stores a vector per token, bloating your storage like a meat wallet stuffed with useless NFTs. The 0.6B version runs on anything that isn’t total scrap, letting you query a heavy index locally. It’s not perfect—ViDoRe Markdown scores are weak—but mixing sizes lets cheap queries access expensive indexes. Perfect for agents that need to read PDFs without burning through your hash manifest budget. Just don’t expect it to fix your bad portfolio choices.