Meta’s 28B encoder turns images, video and text into one search index

Meta’s AI team has posted meta-encoder-v1 on Hugging Face, a roughly 28-billion-parameter model that turns text, images and video into vectors for retrieval. The same vectors can also answer yes/no and multiple-choice questions about a scene.

The page appeared on November 12 under the facebook organization. It is part of a collection called MetaEncoder, which the page describes with the line “One Encoder for Retrieval and Decision Making.” The page carries no paper, no blog post and no license text that we could read, so most of what follows comes from the model card alone.

One vector type for search and for yes/no calls

The model produces embeddings, meaning fixed-length lists of numbers that place similar items close together. According to the card, each embedding has 6,656 dimensions and is L2-normalized, so comparing two items comes down to cosine similarity. The vector is taken from the hidden state at the last position of the prompt, which is the end-of-turn token of the user message, with left padding.

Inputs can be plain strings, images, or dictionaries mixing text, images, video and an optional instruction. Images keep their native resolution, snapped to a 28-pixel grid, up to 4,096 merged tokens. Video is processed frame by frame, and clips longer than 8 frames are sampled down uniformly to 8. Nothing is truncated at inference, though the card says training sequences were capped at 16,384 tokens.

A wrapper class called MetaEncoder exposes four main calls:

• encode returns the embeddings, optionally steered by an instruction.
• retrieve ranks a corpus against a query and returns the top hits.
• decide answers questions about a given state, which can include images or video. Questions can be yes/no, a choice among 2 to 26 options, or a score on ordered levels.
• calibrate fits a temperature on labeled data, so the probabilities attached to each decision can be tuned.

Each decision comes back with a value, probabilities and a confidence figure defined as one minus the normalized entropy. By default the probabilities use a softmax over the cosine scores with a temperature of 0.02, which the card says matches the contrastive training setup. In plain terms, a question and its candidate answers are all embedded, and the closest answer wins.

Who can run it today, and what is missing

The weights are stored in BF16 as Safetensors, and the card says a single 80 GB GPU holds the model. It needs transformers version 5.15 or later, described as the first release to include the muse_glimmer architecture, and the underlying class is named MuseGlimmerModel. A linear embedding head ships as embedding_head.safetensors. Colab and Kaggle notebooks are linked from the page.

The card leaves out several things a buyer would want. It does not describe the training data, name a license, or cite a paper. It links a benchmark image but gives no numeric results in text, so Meta’s performance claims cannot be checked from the page. Anyone planning commercial use has to wait for the license terms. Meta’s most recent Hugging Face release, StereoSplat on November 3, is under a CC BY-NC 4.0 non-commercial license, but the new page says nothing about licensing, so that cannot be assumed to apply here.

The page also does not say how the model relates to Meta’s consumer products or to its other model families. The muse_glimmer tag is the only visible hint at a lineage, and the card does not explain it.

A heavyweight in a field of smaller embedders

Size is the clearest difference from the rivals. Alibaba’s Qwen3-VL-Embedding comes in 2B and 8B versions, with up to 4,096 dimensions, a 32K context length and an Apache 2.0 license, according to its Hugging Face model cards. Google’s Gemini Embedding 2, which maps text, images, video, audio and documents into one space, defaults to 3,072 dimensions and accepts 8,192 input tokens, according to Google’s documentation. Meta’s model is several times larger than the biggest Qwen option and produces much wider vectors, at 6,656 dimensions.

That has a practical cost. A wider vector means larger indexes and slower similarity search at scale, and a 28B encoder is far more expensive to run per document than an 8B one. Whether the extra capacity pays off is exactly what the missing benchmark numbers would show. The Qwen team’s technical report gives its 8B model a score of 77.8 on the MMEB-V2 benchmark, according to a third-party summary page, but the Meta page publishes no figure that can be set against it.

The second distinction is the pitch. Most embedding models are sold for search. Meta’s card also frames the same vectors as a way to make decisions, with built-in calibration and confidence scores. That would let one model sit inside an agent or a monitoring pipeline and answer closed questions about what a camera or a screen shows, without a separate classifier. It is a design claim by Meta, and the page offers no accuracy data to back it.

For now the release reads as a technical drop for researchers and engineers with an 80 GB GPU. Retrieval quality, the license and the data behind it are the three open questions that decide whether it matters beyond Meta.