Meta AI Timeline: Llama 4 to Muse AI and Muse Wonder

Meta AI began in 2013 as Facebook Artificial Intelligence Research, usually called FAIR. It has since grown from a research lab into a company-wide AI program that includes Llama models, computer vision systems, translation research, generative media tools, business agents, and the Meta AI assistant across Facebook, Instagram, WhatsApp, Messenger, smart glasses, and the web.

This timeline tracks the major Meta AI milestones from FAIR’s founding through the latest Meta AI updates in 2026. It focuses on dates, products, research releases, and policy moments that changed how Meta builds or uses artificial intelligence.

Last updated: October 15, 2026.

Meta AI Timeline at a Glance

2013: Facebook created FAIR and hired Yann LeCun to lead its new AI research lab.
2017: PyTorch became one of the most important open source tools linked to Facebook and Meta AI research.
2022: Meta released several pre-Llama AI projects, including OPT-175B, NLLB-200, Make-A-Scene, Make-A-Video, Galactica, and ESMFold.
2023: LLaMA, Llama 2, Segment Anything, ImageBind, I-JEPA, and the Meta AI assistant made Meta a central player in open and consumer AI.
2024-2026: Meta expanded Llama, launched a standalone Meta AI app, and introduced the Muse model family and personal agent. October 2026 brought Muse Wonder, downloadable Muse Spark Open weights, and natural-language controls for Facebook recommendations.

Timeline of Meta AI

Date	Event	Details
2026-10-15	Muse Spark Open	Meta released an open-weight multimodal model for coding, multi-step reasoning, and extended agent tasks. It has 816B total parameters, 42B active parameters, and a 1M-token context window. Inputs include text, images, video, and audio, with text output. Weights are downloadable, and hosted APIs are available through inference partners.
2026-10-13	Your algorithm on Facebook	Facebook began rolling out AI-powered controls for Feed and Reels recommendations. People can add and manage topics from the three-dot menu and explain their interests in natural language after selecting Interested or Not interested. Large language models interpret this feedback, and recommendations adapt within the current session.
2026-10-12	Muse Wonder and Muse Wonder Full	Meta introduced two models for personal and professional agents, specialized knowledge tasks, and coding. Active Memory retrieves relevant context across conversations, while Active Focus supports persistent work through blockers, revised plans, and testing. Muse Wonder launched through Meta Model API and Muse Code. Muse Wonder Full launched for Muse Code subscribers. Integration into Muse is planned, with subscribers first.
2026-10-02	Meta Superintelligence Scaling Framework update	Meta extended its model safety framework to address containment during training and evaluation. Requirements for high-risk reinforcement learning runs include vetted sandboxes, protected logs, and automated monitoring that can halt a run. The update also clarified risk assessments for open-weight releases. Meta announced plans for a board AI committee to review future framework changes and compliance.
2026-10-02	Muse Spark mathematics collaborations	Meta shared six papers from collaborations between mathematicians and Muse Spark 1.1 and 1.2 in Thinking Mode through meta.ai. Five presented answers to previously open research questions. Mathematicians directed the work, and a separate group reviewed it. The papers distinguish human and AI contributions and acknowledge independent solutions from other teams.
2026-09-29	Muse for Small Business	Meta expanded Muse with business skills and connectors for services such as Shopify, Stripe, QuickBooks, Slack, and Canva, plus Facebook and Instagram business accounts. Tasks include sales analysis, campaign drafts, financial reviews, and email follow-up. Publishing, sending messages, and spending require approval. Muse also supports custom connectors.
2026-09-28	Meta Enterprise Platform	Meta announced an enterprise initiative to bring its AI stack to businesses and developers, with an initial focus on Muse, Meta Business Agent, Muse API, and Muse Code. Chirantan Desai joined Meta as Chief Enterprise Platform Officer to lead the effort.
2026-09-23	Muse Realtime Avatar	Meta introduced avatar technology that pairs Muse Realtime Voice with synchronized video for live conversations. It animates portraits, full-body illustrations, animals, and objects from reference media. Meta reported 448-by-768 video at 25 frames per second and approximately 870 milliseconds to the first synchronized response. Generated video carries a Meta Video Seal watermark. The research demos include avatars beyond those available in the Muse app.
2026-09-17	Muse personal AI agent	Muse is now available on Mac.
2026-09-16	Muse Code	Muse Code from Meta is now available natively on Windows. It does not require WSL.
2026-09-08	Muse personal AI agent	Meta launched Muse, a personal AI agent that retains context over time and completes tasks such as managing schedules, shopping, and developing action plans. Each agent runs in a dedicated virtual machine. Meta published a technical security overview alongside the launch.
2026-09-02	Muse Spark 1.3	Meta released Muse Spark 1.3 for Muse Code and Meta Model API. The model is designed for longer agentic workflows, with improved collaboration, complex instruction following, multitasking across messy threads, and awareness of its own limits. Meta also reports stronger resistance to prompt injection and better judgment around consequential actions. In comparisons by Meta engineers, Muse Spark 1.3 used about 20% fewer tool calls and 25% fewer tokens than Muse Spark 1.2.
2026-09-01	Muse Voice Transcribe	Meta introduced Muse Voice Transcribe, the first real-time audio perception model from Meta Superintelligence Labs. It provides streaming automatic speech recognition, speaker diarization for more than 20 speakers, and speech endpoint detection. The model was trained on more than 70 languages, with 25 languages extensively validated, and supports code-switching, context biasing, and audio longer than one hour. It is available through Meta Model API, Meta AI for Mac, and Muse Code.
2026-08-26	Muse Image	Muse Image is now available on Meta Model API and priced for production volumes at $0.01/image. All API Pricing
2026-08-10	Muse Glimmer	Meta released Muse Glimmer, an open-weight 30B-parameter model optimized for always-on local agent workflows. It can run on a Mac or a PC with a single consumer GPU and supports local agents, function calling, coding, and LLM-as-a-judge evaluation. Meta released the model weights under the Apache 2.0 license. Model weights · Resources
2026-08-05	Muse Code	Meta announced Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2. It plans, implements, and validates complex, multi-file changes across large repositories with persistent sub-agents that solve difficult problems faster, more accurately, and with less intervention.
2026-07-09	Muse Spark 1.1 and Meta Model API	Meta released Muse Spark 1.1 with improvements in tool use, computer use, coding, and multimodal understanding. The model became available in Thinking Mode in the Meta AI app and on meta.ai. Meta also launched the public preview of Meta Model API for developers.
2026-07-07	Muse Image and Muse Video	Meta introduced Muse Image and Muse Video, media generation models from Meta Superintelligence Labs for image creation, video generation, and AI effects across Meta apps.
2026-06-03	Meta Business Agent	Meta announced Business Agent, an AI product for businesses that can support customer conversations, personalize responses, and help companies handle more customer interactions across Meta surfaces.
2026-05-13	Incognito Chat	Meta announced Incognito Chat for Meta AI in WhatsApp and the Meta AI app, with private, temporary conversations designed so Meta cannot read the questions or answers.
2026-04-08	Muse Spark	Meta introduced Muse Spark, the first Muse model from Meta Superintelligence Labs. It is a multimodal reasoning model with tool use, visual reasoning, and multi-agent orchestration.
2025-10-01	AI interactions for recommendations and ads	Meta said it would use people’s interactions with AI at Meta to personalize content and ads. This made Meta AI data use and privacy controls part of the company’s AI history.
2025-06-11	Meta AI video editing	Meta launched generative AI video editing in the Meta AI app, the Meta.AI website, and the Edits app, using preset prompts to change short videos.
2025-04-29	Meta AI app	Meta launched the first version of the standalone Meta AI app, built with Llama 4, with voice conversations, a Discover feed, and companion features for AI glasses.
2025-04-29	Llama API	Meta added Llama API as a developer product with API key creation, playgrounds, privacy controls, and access to recent Llama models.
2025-04-05	Llama 4	Meta released Llama 4 Scout and Llama 4 Maverick, its first open-weight natively multimodal Llama models, and previewed the larger Llama 4 Behemoth teacher model.
2024-12	Llama 3.3	Meta released Llama 3.3 70B, before the Llama 4 generation.
2024-10	Movie Gen	Meta presented Movie Gen, a media generation research model for video, audio, and video editing from text prompts.
2024-10	NotebookLlama	Meta’s Llama Recipes project added NotebookLlama, an open source recipe for generating podcast-style audio summaries from documents.
2024-10	Spirit LM	Meta announced Spirit LM, an open source language model designed to handle both speech and text.
2024-10	Lightweight quantized Llama models	Meta released lightweight quantized Llama models built to run on many popular mobile and edge devices.
2024-09-25	Llama 3.2	Meta released Llama 3.2, adding vision models in 11B and 90B sizes and small text-only models in 1B and 3B sizes for edge and mobile use.
2024-08	Sapiens	Meta presented Sapiens, a family of models for human-centric computer vision tasks such as body pose, segmentation, depth, and surface normal estimation.
2024-07-29	AI Studio	Meta launched AI Studio, a place for people and creators to build custom AIs without technical skills.
2024-07-29	Segment Anything Model 2	Meta introduced SAM 2, a unified segmentation model that can identify target objects in images and video.
2024-07-23	Llama 3.1	Meta released Llama 3.1, including 8B, 70B, and 405B models, with support for advanced open model use.
2024-07	Multi-token prediction	Meta shared research on training language models to predict multiple future tokens at once, with code generation gains under the same training budget.
2024-07	JASCO	Meta researchers presented JASCO, a text-to-music generation project that uses audio and symbolic controls such as chords and beats.
2024-07	Vision-language modeling research	Meta researchers contributed an introduction to vision-language modeling, covering how these models connect images, language, and video.
2024-07	Chameleon	Meta released Chameleon, a family of mixed-modal models that can handle text and images as both input and output.
2024-04-18	Llama 3	Meta introduced Llama 3 and positioned it as the strongest openly available Llama generation at the time.
2024-04-18	Meta AI assistant expansion	Meta made Meta AI easier to access across its apps and on the web as part of the Llama 3 launch cycle.
2024-04	Next-generation MTIA	Meta announced its next-generation Meta Training and Inference Accelerator, a custom chip designed for AI workloads.
2024-02	V-JEPA	Meta introduced Video Joint Embedding Predictive Architecture, a video learning model tied to Yann LeCun’s work on more human-like AI systems.
2024-02	AI-generated content labels	Meta said it was working with partners on common technical standards for labeling AI-generated images and later expanded the effort to audio and video.
2024-02	AudioSeal	Meta released AudioSeal, a research system for localized watermarking and detection in AI-generated speech.
2023-12	Seamless Communication	Meta released a family of AI research models for speech and text translation across languages.
2023-11	Audiobox	Meta announced Audiobox, a foundation research model for generating voices and sound effects from voice inputs and natural language prompts.
2023-11	Emu Video and Emu Edit	Meta introduced Emu Video for text-to-video generation and Emu Edit for instruction-based image editing.
2023-10	Brain decoding research	Meta shared research using magnetoencephalography to decode visual representations in the brain at high temporal resolution.
2023-09-27	Meta AI assistant and AI characters	Meta introduced new AI assistants, characters, and creative tools across its apps and devices, bringing Meta AI into consumer products.
2023-07-18	Llama 2	Meta and Microsoft introduced Llama 2, the next generation of Meta’s language model family, with free research and commercial use.
2023-06	I-JEPA	Meta released I-JEPA, an image model based on Yann LeCun’s Joint Embedding Predictive Architecture approach.
2023-05	ImageBind	Meta introduced ImageBind, a model that can bind data from six modalities, including images, text, audio, depth, thermal, and motion data.
2023-04-17	DINOv2	Meta released DINOv2, a self-supervised computer vision method for training high-performance visual models.
2023-04-05	Segment Anything	Meta introduced the Segment Anything project, including SAM and the SA-1B dataset with more than 1 billion masks for image segmentation research.
2023-02-24	LLaMA	Meta released LLaMA, a foundational language model family with 7B, 13B, 33B, and 65B parameter versions for research access.
2022-11-16	Galactica	Meta researchers released Galactica, a large language model for science. Its short-lived public demo became an important cautionary moment because scientific-sounding AI output contained factual errors.
2022-11-01	ESM Metagenomic Atlas	Meta AI shared ESMFold and a database of more than 600 million predicted metagenomic protein structures.
2022-09-29	Make-A-Video	Meta announced Make-A-Video, a text-to-video generation research system that extended generative AI work from images into short video.
2022-07-14	Make-A-Scene	Meta showcased Make-A-Scene, a generative AI research concept that combined text prompts with freeform sketches for more controlled image generation.
2022-07	NLLB-200	Meta AI built NLLB-200, a single translation model covering 200 languages, and open sourced tools tied to the No Language Left Behind project.
2022-05-03	OPT-175B	Meta AI shared OPT-175B, a 175 billion parameter language model released for research access with code, model notes, and smaller baseline models.
2021-10	Facebook becomes Meta	Facebook changed its company name to Meta, and FAIR’s public identity later became part of the Meta AI brand.
2017	PyTorch grows from Facebook AI research	PyTorch became a major open source machine learning framework associated with Facebook AI Research and later the wider Meta AI ecosystem.
2016-09	Partnership on AI	Facebook joined Amazon, Google DeepMind, IBM, and Microsoft in forming the Partnership on AI to discuss best practices and public impact around artificial intelligence.
2013-12	FAIR founded	Facebook created its AI research lab and hired Yann LeCun to lead the effort, setting the foundation for what later became Meta AI.

What Changed Most in Meta AI’s History?

The first phase of Meta AI centered on research: deep learning, computer vision, translation, and open source tools. FAIR helped Meta build long-term AI capacity while publishing research and releasing software such as PyTorch.

The second phase began before Llama, when Meta released large-scale models such as OPT-175B, NLLB-200, Galactica, and ESMFold. These projects showed Meta’s interest in open research, multilingual AI, scientific discovery, and large language models before the widespread adoption of consumer chatbots.

The third phase started in 2023 with LLaMA and accelerated through Llama 2, Llama 3, Llama 4, the Meta AI assistant, and the Meta AI app. Meta AI moved from research papers and model releases into products used inside social apps, smart glasses, business messaging, and creative tools. In 2026, the Muse family extended that product direction into coding agents and personal agents with persistent context. Business connectors and Meta Enterprise Platform expanded the commercial focus, while Muse Wonder added memory retrieval and sustained task execution to the model roadmap. Muse Spark Open followed with downloadable weights for local deployment and customization under the Muse Model License.

Meta AI FAQ

When did Meta AI start?

Meta AI started in 2013, when Facebook created Facebook Artificial Intelligence Research, better known as FAIR. The lab was led by Yann LeCun and later became part of Meta’s main AI organization.

What is the difference between FAIR and Meta AI?

FAIR was the original Facebook AI research lab. Meta AI is the public name for Meta’s AI work, including research, Llama models, the Meta AI assistant, developer tools, generative media systems, and AI features inside Meta products.

What is Muse?

Muse is Meta’s personal AI agent at muse.ai. Its memory and files live in a dedicated virtual machine, where you can inspect, edit, and download them. Muse supports connections to external services, and its settings include an option to exclude your data from model training.

How does Meta evaluate Muse Wonder on long tasks?

Meta uses FocusBench to evaluate persistence and SWE-Marathon to test sustained work on software projects. Muse Wonder and Muse Wonder Full build on larger base models than Muse Spark, and Meta also evaluates their performance on unfamiliar tasks outside the areas targeted during post-training.

When did the Llama timeline begin?

The Llama timeline began on February 24, 2023, when Meta introduced LLaMA as a foundational language model family for research. Llama 2 followed in July 2023, Llama 3 in April 2024, Llama 3.1 and 3.2 later in 2024, Llama 3.3 in December 2024, and Llama 4 on April 5, 2025.

Can Muse Spark Open run locally?

Yes. Meta demonstrated Muse Spark Open through MLX on two M5 Ultra Mac Studios, each with 512 GB of unified memory. That configuration provides 1 TB combined. The demonstration used Muse Code for software changes and video editing tasks entirely on-device.

What license does Muse Spark Open use?

Muse Spark Open uses the Muse Model License, with attribution and usage requirements. Model-as-a-service providers whose aggregate revenue with affiliates exceeds $20 million over any consecutive 12 months must notify Meta and enter a written agreement before offering the model or derivatives through that business. Internal use and other specified commercial uses are exempt from that condition.