---
type: "post"
id: "meta-ai-published-model-weights-on-hugging-face-2c8206"
slug: "meta-ai-published-model-weights-on-hugging-face-2c8206"
title: "Meta AI published model weights on Hugging Face: meta-encoder-v1"
description: "Meta has published MetaEncoder v1 weights on Hugging Face: a 28-billion-parameter model for multimodal embeddings, retrieval and structured decisions."
retrieval_nugget: "Meta has published MetaEncoder v1 weights on Hugging Face: a 28-billion-parameter model for multimodal embeddings, retrieval and structured decisions. Builders can run the supplied Transformers interface locally, but the stated bf16 setup needs one 80 GB GPU, making hardware requirements central to evaluating whether it fits a search or classification workflow."
published_at: "2026-11-12"
updated_at: "2026-11-12"
record_date: "2026-11-12"
date_kind: "published_at"
topics: ["retrieval"]
entities: ["Meta","MetaEncoder v1","Hugging Face","MuseGlimmer"]
related_patterns: ["agent-ready-software-exposes-capabilities"]
novelty_label: "new_fact"
what_changed: "Meta AI published meta-encoder-v1 weights on Hugging Face, a distinct release from the other models covered."
prior_refs: ["https://newruntime.com/posts/bytedance-seed-published-model-weights-on-hugging-face-8029c4/"]
source_url: "https://huggingface.co/facebook/meta-encoder-v1"
source_urls: ["https://huggingface.co/facebook/meta-encoder-v1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:meta-ai-published-model-weights-on-hugging-face-2c8206"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/meta-ai-published-model-weights-on-hugging-face-2c8206/","markdown":"https://newruntime.com/posts/meta-ai-published-model-weights-on-hugging-face-2c8206.md","json":"https://newruntime.com/posts/meta-ai-published-model-weights-on-hugging-face-2c8206.json"}
markdown_url: "https://newruntime.com/posts/meta-ai-published-model-weights-on-hugging-face-2c8206.md"
---

# Meta AI published model weights on Hugging Face: meta-encoder-v1

## Retrieval answer

Meta has published MetaEncoder v1 weights on Hugging Face: a 28-billion-parameter model for multimodal embeddings, retrieval and structured decisions. Builders can run the supplied Transformers interface locally, but the stated bf16 setup needs one 80 GB GPU, making hardware requirements central to evaluating whether it fits a search or classification workflow.

## Summary
Meta has published MetaEncoder v1 weights on Hugging Face: a 28-billion-parameter model for multimodal embeddings, retrieval and structured decisions. Builders can run the supplied Transformers interface locally, but the stated bf16 setup needs one 80 GB GPU, making hardware requirements central to evaluating whether it fits a search or classification workflow.

## What happened
Meta's Hugging Face model card describes MetaEncoder v1 as a feature-extraction model, with tags for multimodal embedding, image-text retrieval, video retrieval and document retrieval. The repository contains sharded BF16 safetensors totaling 28,431,795,200 parameters and 56,863,902,482 bytes. The card provides a Transformers pipeline example and direct loading through `AutoProcessor` and `AutoModel`; its wrapper requires `transformers>=5.15`.

The wrapper exposes three calls. `encode` embeds inputs for storage, clustering or comparison and returns an `(n, 6656)` float32 tensor with L2 normalization. Inputs can be strings, PIL images or structured items containing text, images or video. Callers can supply task instructions; the card says the same input can produce different embeddings under different instructions, and query and document instructions usually differ. The wrapper formats prompts in the layout used in training, so callers do not have to write that formatting themselves.

`retrieve` ranks a corpus against a query and returns hits in best-first order, including the corpus position, cosine similarity and item. A corpus may be supplied as items or as a tensor produced by `encode`, allowing a large corpus to be encoded once and searched repeatedly. `decide` accepts a state and named questions, with question types for yes/no, choice and score. It returns answers with probabilities, confidence and scores; the card describes this as retrieval over candidate options rather than a generative response.

The model accepts images and video alongside text. Images retain native resolution, snapped to a 28-pixel grid, up to 4,096 merged tokens. Video is encoded frame by frame; inputs with more than eight frames are uniformly subsampled to eight. Training sequences were at most 16,384 tokens. Meta's card says one 80 GB GPU can hold the model in BF16. It also shows a direct loading path without the wrapper, including the embedding head and the MuseGlimmer model. Hugging Face currently lists no Inference Provider deployment for the repository.

## What the archive adds
New Runtime's earlier coverage documented other distinct Hugging Face weight releases: [Kimi K3.1 Pro](https://newruntime.com/posts/moonshot-ai-released-full-open-weights-for-kimi-8c09c3/) on 2026-11-04 and [ByteDance Seed's Cola2](https://newruntime.com/posts/bytedance-seed-published-model-weights-on-hugging-face-8029c4/) on 2026-11-06. MetaEncoder is a new event, not a continuation of either model release; the shared publishing venue does not make their capabilities interchangeable.

## Trend signal
This is a weak fit for New Runtime's [agent-ready software exposes capabilities](https://newruntime.com/patterns/agent-ready-software-exposes-capabilities/) pattern: the model card presents bounded calls for embedding, retrieval and decision-making rather than only a chat interface. That is an interface choice, not evidence that the model has been integrated into an agent product or safely permissioned for external agents. The release alone does not establish a broader pattern shift.

## Why it matters
For teams building multimodal search, the distinction between an embedding model and a conversational model changes how the system is assembled: encode content into a reusable corpus representation, retrieve candidates for a query, then apply a downstream decision or review step. Reusing an encoded tensor may avoid repeating corpus encoding for every query, although the card does not quantify the savings or retrieval quality.

The hardware bar is a practical constraint. A stated 80 GB GPU requirement for BF16 may rule out small local deployments, and no hosted Inference Provider is listed on the page. Teams should treat this as an evaluation candidate rather than a drop-in service: test retrieval on their own content, compare instruction choices for queries and documents, and account for indexing hardware and latency before committing.

## What happens next
- Hugging Face will add at least one Inference Provider deployment for MetaEncoder v1 within 6 months (35%). Watch for a provider entry on the model page.

## Limits and uncertainty
The model card describes interfaces and hardware requirements, but the page excerpt does not establish benchmark quality, latency, throughput, or comparative performance on image, video or document retrieval. The listed parameter count and file size do not indicate the cost of indexing a production corpus. No hosted inference provider is currently shown, and the card does not establish a license interpretation or deployment terms beyond what appears on the repository page. Independent evaluation on representative workloads remains necessary.

## Skill to build now
Build a small retrieval evaluation set from your own image or document corpus: write representative queries, label the relevant items, and compare results under distinct query and corpus instructions. Use `encode` to create corpus embeddings once, then run `retrieve` and record recall at your chosen top-k alongside GPU memory and indexing time. This gives a concrete basis for deciding whether the 80 GB BF16 setup is worthwhile before scaling.

## Reader takeaway
MetaEncoder v1 is a downloadable multimodal retrieval and decision model, not a chat model. Its encode-once/search-repeatedly interface is useful to investigate, but the 28B BF16 weights and stated 80 GB GPU requirement make workload-specific quality and hardware testing the first step.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Retrieval](https://newruntime.com/topics/retrieval.md): Explore the retrieval topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)