Meta

Meta Platforms, Inc. is the company behind Facebook, Instagram, Messenger and WhatsApp, and a developer of AI infrastructure, models and applications. PyTorch, Llama and Segment Anything give engineers tools for building systems beyond its own services. Its newer Muse agent combines persistent work on a user's behalf with an architecture that separates credentials and execution permissions from the model's control. Advertising remains its dominant revenue source: Meta reported $59.363 billion in advertising revenue within $60.801 billion of total revenue for the second quarter of 2026.[23][3][5][8][14]

These AI products grew within an organization whose social apps and hardware investments predate the Meta name. Facebook launched in 2004; the company introduced the Meta brand in October 2021, while its FAIR research program began in late 2013.[1][2]

The company was incorporated in Delaware in July 2004 and completed its initial public offering in May 2012. Facebook sold 180 million company shares at $38 per share and received $6.76 billion in net proceeds. Selling stockholders separately sold 241,233,615 shares; their sales did not provide capital to Facebook.[17][18]

Facebook announced an agreement to acquire Instagram in April 2012 and welcomed its team that September. Mike Schroepfer described a plan to grow Instagram independently while supporting it with Facebook's engineering team and infrastructure. Facebook later completed the acquisition of WhatsApp on October 6, 2014. Its 2014 filing described redirecting messaging users to standalone Messenger and growing Instagram and WhatsApp despite limited or absent near-term monetization.[19][20][21][17]

The Oculus VR acquisition, completed in July 2014, extended Facebook into virtual reality. At the time, the company acknowledged that it lacked prior experience with consumer hardware and VR technology. On October 28, 2021, CEO Mark Zuckerberg introduced Meta as the brand bringing together the company's existing apps and technologies, with a focus on the metaverse and immersive social experiences. Corporate structure remained unchanged; the announcement introduced financial reporting through Family of Apps and Reality Labs beginning with fourth-quarter 2021 results.[17][1]

Meta's AI research investment began before that brand change. FAIR launched in late 2013. In its ten-year retrospective, Joelle Pineau recalls Zuckerberg traveling to NeurIPS to recruit researchers and partnering with Yann LeCun to assemble the initial team. The research portfolio developed across object detection, segmentation, translation and speech: Mask R-CNN followed in 2017, and later work included No Language Left Behind and speech technology covering more than 1,000 languages.[2]

PyTorch, incubated by Meta's AI team and developed with a wider contributor community, moved to a foundation under the Linux Foundation in September 2022. The foundation announcement credits Soumith Chintala as a co-creator and describes a neutral home with communal governance. The transition gave a tool originating inside Meta an institutional structure intended to serve users and contributors across organizations.[3]

Announced in November 2022, CICERO coupled a controllable dialogue model to a planning engine for Diplomacy. The engine predicted other players' likely moves from the board and conversation history, then selected plans that conditioned its negotiation messages. Researchers associated with the project included Mike Lewis, Noam Brown and Anton Bakhtin. In a published gameplay example, CICERO negotiated peace with France by asking for Munich's return and agreeing to divide English supply centers.[9][10]

Meta reported that across 40 games against human players, CICERO achieved more than twice the average human score and ranked in the top 10% of participants who played more than one game. The demonstrated domain was Diplomacy; proposed applications to assistants that teach skills or negotiate for people remained prospective.[10][9]

The February 2023 LLaMA paper, by Hugo Touvron and colleagues, introduced language models ranging from 7 billion to 65 billion parameters. Building on earlier scaling research, it emphasized the cost of repeatedly running a model: a smaller model trained on more tokens can require additional training effort while becoming cheaper to serve afterward. The authors used publicly available datasets, which does not mean all the material was in the public domain. Their evaluations also documented incorrect-answer, toxicity and bias risks.[4]

Llama 2's July 18, 2023 release expanded access to research and commercial use, supplying pretrained and conversationally fine-tuned weights with starting code. Microsoft was the preferred partner, with Azure distribution and Windows optimization; AWS and Hugging Face also offered distribution. Downloadable weights allowed developers to modify the model and choose where to run it, subject to the Community License and acceptable-use conditions. Free model licensing did not remove hosting or inference costs.[5][6]ⁿ¹

In the published Latent Space launch discussion, Simon Willison called Llama 2 “openly licensed, not open source.” Nathan Lambert, whose Hugging Face team collaborated on launch support and received early details, distinguished the paper's extensive methodological explanation from its incomplete dataset disclosure. These were separate dimensions of access: developers could obtain and adapt weights without receiving a complete account of the training data.[16][6]

Meta's 2023 research releases also included Segment Anything. The original SAM project, credited to Alexander Kirillov, Eric Mintun, Nikhila Ravi and a wider team, turns points or boxes into object masks specifying the selected pixels. Its repository supplies prompted prediction and automatic mask generation. The lightweight mask decoder can be exported for ONNX Runtime, including browser use, while image encoding remains part of the workflow: encode the image, then use prompts to select regions. The reported 11 million images and 1.1 billion masks describe its training dataset, whose downloads have a separate research license.[2][8]

Alongside the July 2024 Llama 3.1 release, with 405B, 70B and 8B models, Zuckerberg explained his strategy in “Open Source AI Is the Path Forward”. He argued that Llama needed an ecosystem of tools, efficiency improvements, silicon optimizations and integrations that would also benefit Meta. He linked his preference for open ecosystems partly to his experience building services within Apple's platform rules. In that essay, he described selling model access as outside Meta's business model and presented broader adoption as a way to reduce dependence on competing closed platforms.[7]

With Llama 4 Scout and Maverick in April 2025, Meta released downloadable multimodal models using mixture of experts. Early fusion integrated text and vision tokens in a unified backbone. Maverick used 17 billion active parameters out of 400 billion total; in its expert layers, each token passed through a shared expert and one selected routed expert. This reduced active computation while leaving the full parameter set to be stored. Scout's single-H100 deployment claim required Int4 quantization; Maverick's concerned an H100 host. Behemoth was previewed as a teacher model still in training.[11]

Meta Superintelligence Labs introduced Muse Spark on April 8, 2026 as the first Muse-family model following a ground-up overhaul of its AI efforts. It supported multimodal reasoning, tool use and multi-agent orchestration. Its Contemplating mode used agents reasoning in parallel. Launch access included meta.ai, the Meta AI app and a private API preview for selected users. Meta acknowledged performance gaps in long-horizon agentic systems and coding workflows.[12]

The June 2026 Meta Glasses launch connected the model to consumer hardware through a partnership with EssilorLuxottica. The line started at $299, with prescription-lens compatibility, open-ear audio, hands-free photo and video capture, and a dedicated action button. Meta AI powered by Muse Spark was available at launch in the US and Canada, according to the announcement's July update.[24]

On September 8, 2026, Meta began rolling out Muse, a personal agent powered by Muse Spark, in the US on iOS, Android and muse.ai, with interaction through its app or WhatsApp. Its dedicated virtual computer and browser supported tasks that continued after the app closed, including forms, travel arrangements, email and shopping. Stripe Link supplied checkout integration using single-use payment credentials. The launch offered free usage and subscription plans for people who wanted to do more.[13]

In “How We Built Safety Into Muse”, Tarek Sheasha recalls internal use beginning in early 2026, when unattended access to inboxes, calendars and a shell “didn’t always work out as planned”. The launch architecture separated an unprivileged agent runtime from host-side services holding credentials, connector logic, permissions and durable state. Its design assumed that the agent could make mistakes or encounter attacks through material it read.[14]

• Policy checks: Sentinel governed connector actions and network egress, allowing requests, denying them or asking the user for approval under the relevant policy.[14]
• Credential handling: Runtime code used surrogate tokens; Sentinel inserted real credentials at the network boundary after authorizing the concrete request.[14]
• Scoped approvals: Client dialogs sent the user's decision directly to Sentinel. Permission grants were bound to particular connectors or destinations and use cases, allowing subsequent matching operations without a new dialog each time.[14]

Purchases required stricter approval. The documented controls requested human approval with exact purchase details at every checkout using stored payment credentials and whenever a single-use payment credential was issued. The browser subagent operated through a separate broker using accessibility snapshots, with no arbitrary page JavaScript execution; agent activity paused during user takeover or secure credential entry. These were Meta's described defenses, with prompt injection and mistakes still acknowledged as possible.[14]

Muse's launch isolation did not prevent Meta from accessing data when necessary to support, secure or operate the service. Sanitized conversations, tool calls and subagent trajectories were used for model training by default, with an opt-out. Meta said Muse conversations and VM data were not shared with its ad systems, while browser activity could indirectly influence advertising. Muse Confidential VM, intended to cryptographically prevent provider access to VM contents, was planned for later in 2026.[14][13]

Meta reported $200.966 billion in full-year 2025 revenue and $72.22 billion in capital expenditures, including finance-lease principal payments. For June 2026, it estimated 3.60 billion family daily active people on average—a measure across its app family, rather than Facebook alone or adoption of Meta AI and Muse. Second-quarter 2026 capital expenditures including finance-lease principal were $31.08 billion. These spending figures were company-wide, without an AI-only allocation.[22][23]

Reality Labs, covering VR and AR consumer hardware, software and content, reported $431 million in revenue and a $4.619 billion operating loss for the quarter ended June 30, 2026. Meta as a whole reported $18.775 billion in operating income for that quarter. The segment loss therefore describes the economics of that business unit, rather than a company-wide loss.[23]

A July 28, 2026 announcement described a venture with BlackRock to develop and own an El Paso data center campus. BlackRock-managed funds were to hold 80% and Meta 20%, with commitments to fund their respective shares of approximately $14 billion in development costs for buildings and long-lived power, cooling and connectivity infrastructure. Meta would be the initial sole occupant and provide construction and property management services. The venture expected to begin bringing 1 GW of capacity online in 2028; closing was still prospective at announcement.[15]

The announced financing structure also included leases with a four-year initial term and extension options allowing a potential 20-year term. Meta would provide residual value guarantees with an aggregate threshold of approximately $13 billion that decreased over time, potentially covering a shortfall in property value under specified conditions. The venture combined outside capital with continuing Meta lease and guarantee obligations.[15]