Meta Releases MetaEncoder: A Multimodal Bi-Encoder for Natural Language Retrieval and Decision Making

·2026.11.12 18:28

Key point

Built by contrastively fine-tuning meta-models/Muse-Glimmer-30B, the model supports arbitrarily large candidate sets through ANN indexing.

Meta introduces MetaEncoder, a multimodal system using a single encoder with a natural language interface for instruction-following encoding.
The model scores candidates by task specification, accepting free-form text instructions, queries, and criteria alongside image and video side information.
Candidates and tasks are encoded separately into cacheable representations, allowing matching via inner product over an ANN index to serve millions of candidates without re-running the model.
The architecture is derived from contrastively fine-tuning meta-models/Muse-Glimmer-30B.
Weights are available on Hugging Face as meta-encoder-v1 (improved version) and meta-encoder (Apache-2.0 licensed).

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.