Descriptions:
Sam Witteveen covers the Qwen3-VL-Embedding models—Alibaba’s new multimodal embedding series in 2B and 8B parameter sizes—which place text, images, and video frames into a unified vector space for cross-modal semantic search. The 8B model currently holds the #1 position on the MMEB (Massive Multimodal Embedding Benchmark) leaderboard, with the 2B variant at #5. Both support a 32K token context window.
The video explains the full architecture: a bi-encoder model for fast large-scale recall, paired with a cross-encoder reranker for precision re-scoring of top candidates. A standout technical feature is Matryoshka Representation Learning (MRL), which allows developers to use truncated embedding dimensions—for example, just the first 1,024 values of a 4,096-dimensional vector—to trade off search latency against accuracy at query time, without re-embedding the corpus.
Practical use cases covered include visual document search (embedding PDF pages as images to capture chart and diagram content that traditional OCR discards), e-commerce product search with combined image and text queries, and video surveillance frame retrieval using reference images or natural language descriptions. Witteveen includes a Google Colab walkthrough using the 2B model on a T4 GPU, demonstrating both the embedding API and the reranker API with real examples—making it directly usable for developers building multimodal RAG pipelines with open-weight models.
📺 Source: Sam Witteveen · Published January 15, 2026
🏷️ Format: Tutorial Demo
