Text embedding models are widely used in RAG, semantic search, and other scenarios, but high-quality embeddings have long faced an "impossible triangle": compute cost, language coverage, and model transparency are hard to combine. The ICML 2026 paper ML-Embed attempts to break this deadlock.
ML-Embed proposes a three-dimensional Matryoshka learning framework (3D-ML), optimizing across three dimensions of the model lifecycle: MRL (Matryoshka Representation Learning) reduces storage overhead, MLL (Matryoshka Layer Learning) supports on-demand depth adjustment at inference, and MEL (Matryoshka Embedding Learning) improves parameter efficiency. The model has parameter counts from 140M to 8B, evaluated across 430 tasks, setting 9 new records across 17 MTEB benchmark subsets — with particularly impressive performance on low-resource languages.
What's even more noteworthy is the team's choice to fully open-source the model, data, and code. Against the backdrop of embedding services widely relying on closed-source APIs, this provides a low-cost onramp for academic research and small-to-mid developers.
From an engineering perspective, ML-Embed's three-layer decoupling design is worth learning from — storage, inference, and parameter efficiency are optimized separately, and the feasibility of edge deployment scenarios is significantly improved. How to maintain multilingual coverage while controlling inference latency remains a key issue for follow-up research.