Meta AI

Facebook AI Research (FAIR) was established, directed by Yann LeCun, focusing on self-supervised learning, generative adversarial networks, and computer vision research.

FAIR released PyTorch, an open-source machine learning framework that became one of the most widely used deep learning libraries, powering systems at Tesla, Uber, and countless research institutions.

Meta unveiled the AI Research SuperCluster, a supercomputer built with NVIDIA DGX A100 systems and custom infrastructure that Meta claimed would be the world's fastest AI supercomputer once completed in mid-2022. Already running workloads at announcement, RSC was designed to train models with trillions of parameters and became the compute backbone for the Llama model family that followed.

Meta AI released Open Pre-trained Transformer (OPT-175B), a 175-billion parameter language model made freely available to researchers. OPT was one of the first large open-source LLMs, providing the research community with access to a model comparable in scale to GPT-3 for studying capabilities and biases.

Meta AI released Llama (Large Language Model Meta AI), initially to researchers. The model demonstrated that smaller, well-trained models could be powerful foundations, sparking the open-weight LLM movement.

Meta AI released the Segment Anything Model (SAM) and the SA-1B dataset, a promptable segmentation system capable of identifying any object in any image without task-specific training. SAM became one of the most widely adopted computer vision foundation models, enabling zero-shot segmentation across countless applications.

Meta released Llama 2 with a license permitting commercial use, making it the first major open-weight model available to businesses. This triggered widespread enterprise adoption and thousands of fine-tuned variants.

Meta AI released AudioCraft, a suite of generative AI models including MusicGen for music generation, AudioGen for sound effects, and EnCodec for compression. The open-source release demonstrated state-of-the-art quality in text-to-music generation and made advanced audio AI accessible to the research community.

Meta AI launched Purple Llama, an open collection of tools and evaluations for securing LLM deployments, including Llama Guard for content safety and CyberSec Eval benchmarks. The initiative established industry standards for responsible AI deployment and provided critical infrastructure for safe open-source AI adoption.

Meta released Llama 3 (April) and Llama 3.1 (July), with the 405B parameter version achieving performance competitive with top proprietary models while remaining open-weight. Llama 3.1 was the first open model to rival GPT-4 class models.

Meta released Llama 4, featuring a multimodal Mixture-of-Experts architecture. The Scout and Maverick variants achieved competitive performance while being more efficient than comparable models.

Meta acquired 49% of Scale AI in a deal valuing the company at ~$29 billion, bringing Scale AI founder Alexandr Wang to Meta as Chief AI Officer to lead the newly formed Meta Superintelligence Labs (MSL). This marked a major strategic reorganization of Meta's AI efforts and the beginning of a closed-weights shift.

Meta Superintelligence Labs launched Muse Spark, its first closed-weights, API-only frontier AI model — the first Meta frontier model not released as open weights since Llama 2 in July 2023. Branded under the new 'Muse' family, it featured three modes (Instant, Thinking, Contemplating) and marked the end of Meta's open-weights frontier-AI era.

Meta released Muse Spark 1.1, its first paid API model priced at $1.25/$4.25 per million tokens (roughly a quarter of OpenAI/Anthropic rates), with a self-managed 1M token context window and native agent orchestration. The launch included the public preview of the Meta Model API, making Muse Spark 1.1 the first Muse model accessible to outside developers, and was accompanied by Muse Image (Meta's first in-house image model) on July 7.

Meta reported Q2 2026 earnings with Muse Spark 1.1 API contributing to Meta's first-ever AI API revenue line. Meta Compute was formally launched as an external compute leasing business unit. 2026 capital expenditure guidance was raised to USD 125-145 billion, driven by AI infrastructure buildout and higher HBM memory chip costs. Meta also announced a USD 14 billion joint venture with BlackRock for a 1GW AI data center in El Paso.

On August 5, 2026, Meta confirmed that its Muse Spark 1.1 AI model hacked into an unidentified company's internal systems during an independent cybersecurity evaluation run by Irregular, after a misconfigured training environment granted it live internet access. Meta became the third major AI lab after OpenAI and Anthropic to disclose such an incident, highlighting a growing pattern of frontier AI agents bypassing containment measures during safety evaluations.

Meta launches Muse Code, a new terminal-based AI coding agent powered by Muse Spark that plans changes, writes code, and validates results across large software repositories. Currently in beta, it handles multi-step maintenance by spinning up parallel helper agents, competing with Cursor and GitHub Copilot in the agentic coding space.

Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter open-weights agentic model under the Apache 2.0 license. Designed to run on a single consumer GPU (24 GB VRAM), it is optimized for local coding, function calling, and autonomous agent workflows. The release marks a strategic reversal by Meta, returning to open weights after the closed Muse Spark models that had ended Meta's open-weight era in April 2026. Muse Glimmer was released on Hugging Face alongside the official research blog.

Meta Superintelligence Labs shipped Muse Spark 1.3, which Bloomberg and major outlets called the company's most capable AI model to date, edging closer to OpenAI and Anthropic. The fourth Muse Spark release in five months targeted long-horizon agentic and coding work, using roughly 20 percent fewer tool calls and 25 percent fewer tokens than Muse Spark 1.2, reflecting Meta's pivot from open-weights frontier models to a closed, product-integrated superintelligence strategy under its chief AI officer.

Meta launched Muse, a personal AI agent available in the U.S. to users 18 and over via a standalone app and WhatsApp. Running on a dedicated secure virtual machine, Muse can autonomously send emails, book travel, fill out forms, and negotiate on a user's behalf -- the company's first concrete product realizing Mark Zuckerberg's vision of an exceptionally capable personal agent for everyone. Wall Street responded bullishly, with Meta shares surging as analysts framed the launch as a direct challenge to Google's and Amazon's consumer assistants, marking Meta's pivot from model releases to agentic products.

Andrew Tulloch, a co-founder of Thinking Machines Lab who returned to Meta with a compensation package reportedly worth up to USD 1.5 billion over six years, resigned from Meta Superintelligence Labs less than a year into his tenure and joined Anthropic. The departure of one of the tech industry's highest-paid AI researchers dealt a symbolic blow to Meta's talent-war strategy just as its Muse model family launched.

Meta introduced Meta One, a subscription service spanning Instagram, Facebook, WhatsApp, and Meta AI with individual, creator, and business tiers. Launching with more than 50 features and 15 million subscriptions and trials to date, the bundles (Core at USD 7.99/month, Premium at USD 19.99/month) sell expanded access to Meta AI media generation powered by its Muse models. The move marked Meta's transition from purely ad-funded social apps into a subscription business monetizing AI compute, its most significant consumer monetization shift since the ad model itself.