Meta releases MetaEncoder: one bi-encoder for retrieval and decision-making via natural language
pmttyji · reddit · 2026-11-12
Meta released MetaEncoder, a multimodal single-encoder system with a natural language interface for instruction-following scoring: describe a task and candidates in free-form text (instructions, queries, criteria, with images/video as side info) and the model scores candidates against the task.
- Task and candidates are encoded separately with mutually grounding prompt instructions, yielding cacheable representations;
- Matching is an inner product over representations, so an ANN index can serve millions of candidates without re-running the model;
- Built by contrastively fine-tuning Muse-Glimmer-30B, described in the paper Exploring the Limit of Bi-Encoders for Multimodal System One Decision Making with Natural Language Interface (arXiv:2610.11316);
- Models are open on Hugging Face (meta-encoder and meta-encoder-v1, the former Apache-2.0), with code on GitHub.
More from Models
- LLMs play Among Us against each other; Haiku 5.5 kills then instantly self-reports — Rasmic · 2026-11-12
- Kimi K3.1-Pro abliterated variant hits Hugging Face, claiming 95% refusal removal — zebird0 · 2026-11-12
- Possible eval leak: all frontier models fail a GPQA Diamond question whose official answer seems wrong — qberthet · 2026-11-12
- Engineers Using LLMs Should Understand Tokens, Context Windows and Hallucinations — _jaydeepkarale · 2026-11-12
- Math Benchmark Sheet Shows 90.43% Top Accuracy at Just $1.21 per Run — teortaxesTex · 2026-11-12
- Open model Inkling 2 Small takes top open spot on MathArena's BrokenArXiv — teortaxesTex · 2026-11-12