Meta AI Releases 2024: An Artifact-Level Map

Meta AI Release Map: CoTracker3, Movie Gen, Spirit LM, and Sparsh

CoTracker3, Movie Gen, Spirit LM, and Sparsh were all public Meta research projects in 2024. They were not four equivalent product launches. The right comparison starts with the artifact, access path, license, evidence, and present state of each project.

This map reflects sources checked on July 28, 2026. It preserves the 2024 historical record while showing later changes separately.

flowchart TD
    A["Announcement"] --> B{"What is public?"}
    B --> C["Paper or project page"]
    B --> D["Code and configuration"]
    B --> E["Weights or checkpoints"]
    B --> F["Dataset or data information"]
    B --> G["Demo, pilot, API, or product"]
    C --> H["Read the evidence"]
    D --> I["Read file-level licenses"]
    E --> I
    F --> I
    G --> J["Check access and terms today"]
    H --> K["Run only a task-specific test"]
    I --> K
    J --> K

The four-project view

The table is intentionally blunt. A paper tells you that a method and result were disclosed. A repository tells you that some implementation material is public. Neither tells you that the system is supported, lawful for your use, reliable on your data, or ready for production.

Those questions require a separate task-specific review.

CoTracker3 offered a path from paper to experiment

The CoTracker3 paper describes a point tracker trained with pseudo-labels on real video, alongside synthetic training data. The official repository exposes code and checkpoints. The project page provides selected visual examples.

That combination gives a technically capable reader something to inspect and run. It still leaves deployment questions. A team must test its own camera motion, occlusion, blur, object motion, frame rate, resolution, compute budget, and failure cost.

E046 owns the detailed record. [[How CoTracker3 Works]] explains the method, while [[How to Evaluate a Point Tracking Model]] turns the research into a bounded test.

Movie Gen published evidence without unrestricted model access

Meta's Movie Gen research page presents four capability groups: generating video from text, producing personalized video from an image, editing video through instructions, and producing sound effects or music.

The publication page reports a 30 billion parameter video model that generated up to 16 seconds at 16 frames per second and 1080p. The research record also included selected videos and a creator feedback program with filmmakers.

Those are meaningful research artifacts. They are not equivalent to public code, downloadable weights, or an unrestricted hosted service. Later Meta video features help show a research-to-product path, but they should not be renamed as access to the original model.

Spirit LM released weights under a research boundary

The Spirit LM paper describes a 7 billion parameter language model extended to speech. Spirit LM Base uses phonetic speech units. Spirit LM Expressive adds pitch and style information.

The repository contains weights, inference code, evaluation scripts, and a model card. The controlling license permits covered use for noncommercial research and imposes an acceptable-use policy.

That is more access than a paper alone. It is less freedom than a permissively licensed model intended for any purpose. Calling it simply open source would hide the most important operational fact.

Sparsh made tactile research reusable, then became archived

Sparsh explored self-supervised pretraining across several vision-based tactile sensors. The publication and repository exposed enough material to inspect the representation approach and downstream tasks.

The repository license is noncommercial. Dataset access and sensor hardware have their own terms and practical constraints. The repository was archived on April 1, 2026, which makes maintenance risk visible to a new evaluator.

E045 preserves the technical distinctions in [[How Sparsh Learns Touch Representations]] and [[How to Read the TacBench Results]].

A release taxonomy that survives the next announcement

The safest description names the thing. A public paper is a public paper. Source-available code is code a reader can inspect under stated terms. Downloadable weights are model parameters with a separate license. A research demo is a selected example. A pilot gives a limited group access. A product has current users, terms, support expectations, and an operating surface.

One project can occupy several categories. That is normal. Trouble begins when the broadest category is used to imply every other one.

The related [[What Open Source Means for an AI Release]] provides an artifact test for openness. [[How to Evaluate an AI Research Release]] provides the next step when one of these projects appears relevant to a real workflow.

Editorial note

This dated research reference was developed with AI assistance from the recovered E044 captions and the linked primary papers, repositories, licenses, project pages, announcements, and later product records. Dalton Anderson remains the author. Technical, source, license, current-state, rights, and founder review are mandatory before publication. Publication is not authorized.