Meta
AI released Llama 4 in April 2025, a family of multimodal large language
models that the company claimed outperformed comparable models from OpenAI
and Google on several standard benchmarks. The release continued Meta’s
strategy of open-weight model publication, making the model weights available
for download and use under a licence that permits commercial deployment with
limited restrictions. Within 72 hours of release, Llama 4 had been downloaded
over 100,000 times by developers globally, integrated into dozens of
third-party applications, and subjected to the kind of independent capability
testing and red-teaming that closed models from OpenAI and Anthropic do not
receive at equivalent scale. Meta’s AI laboratories are producing some of the
most capable and most discussed AI systems in the world, and understanding
what they are doing, how they are structured, and what their strategic
intentions are is increasingly important for anyone seeking to understand the
AI landscape.
Meta’s AI research enterprise is structured across two primary
organisations with different but related mandates. FAIR (Fundamental AI
Research), founded in 2013 under Yann LeCun, focuses on long-term
foundational research and has been responsible for some of the most cited
papers in deep learning and computer vision. The GenAI team, which operates
with a more product-oriented mandate, is responsible for the Llama model
family, the Meta AI assistant integrated across Facebook, Instagram,
WhatsApp, and Ray-Ban smart glasses, and the AI features embedded throughout
Meta’s consumer products. The relationship between the two organisations, and
the tension between academic research culture and product development
pressure, is a recurring theme in accounts from current and former Meta AI
researchers.
The Open-Weight Strategy
Meta’s decision to release model weights openly, rather than
through an API with usage restrictions as OpenAI and Google do, is the most
strategically distinctive aspect of its AI approach. The company frames this
as a commitment to open science and to preventing AI capability being
concentrated in a small number of proprietary systems. Critics, including
several AI safety researchers and some of its competitors, argue that the
strategy is primarily a competitive tactic: by making powerful models freely
available, Meta commoditises the base model layer of the AI stack, reducing
the competitive advantage of proprietary model providers while positioning
Meta’s own consumer platforms, which are closed and ad-supported, as the
primary interface through which the value of AI reaches end
users.
The safety implications of open-weight model release are actively
debated. Arguments in favour include the independent auditability that open
weights enable, the ability for safety researchers to examine model behaviour
without depending on company disclosure, and the distributed development of
safety improvements that open models facilitate. Arguments against include
the ability of bad actors to remove safety fine-tuning from released weights,
the difficulty of preventing misuse of a model once weights are publicly available,
and the risk that releasing increasingly capable open models accelerates
proliferation of powerful AI systems without equivalent proliferation of
safety infrastructure. The UK
AI Safety Institute has engaged with Meta on pre-release evaluation
of Llama models, representing one of the first instances of a government
safety body having access to a major open-weight model before public
release.
Meta’s Multimodal Ambitions
Llama 4 and its successors are multimodal, meaning they can
process and generate text, images, and audio in an integrated way. This
capability is central to Meta’s product vision, in which AI assistants
embedded in smart glasses, augmented reality headsets, and social media
platforms need to understand and respond to the full sensory environment of
their users. The Ray-Ban Meta smart glasses, which integrate the Meta AI
assistant with a camera and audio interface, have sold in significant volumes
and represent the most commercially successful consumer AI hardware product
outside the smartphone market.
The convergence of Meta’s multimodal AI capability with its vast
social media data assets and its consumer hardware ambitions creates a
competitive position that differs qualitatively from pure AI model providers.
Meta has access to behavioural data about billions of users, distribution
infrastructure that reaches more people than any other platform, and an
increasingly capable AI system that can personalise interactions in ways that
no competitor can match at equivalent scale. Whether this constitutes a
competitive advantage that benefits users through better AI experiences, or a
data and distribution monopoly that needs regulatory scrutiny, is a question
that competition authorities in the EU, UK, and US are beginning to examine
more seriously.
What This Means for You
If you use Facebook, Instagram, or WhatsApp, Meta AI is already
integrated into your experience in ways that are expanding monthly. The AI
assistant that answers questions in the search bar, generates images in
messaging, and summarises content in your feed is powered by the same Llama
model family that researchers are downloading and studying worldwide.
Understanding that the AI features in consumer social media products are
connected to a larger research enterprise with specific strategic goals, and
that the open-weight model releases that researchers celebrate are part of a
competitive strategy as well as a commitment to open science, is part of
being an informed user of these platforms. The social and psychological
implications of AI assistants integrated into the world’s most used social
media platforms are also beginning to attract research attention. Billions of
people interacting daily with Meta AI through Facebook and WhatsApp are
having their information diet, social connections, and self-perception shaped
by AI systems optimised for engagement rather than wellbeing. The evidence
that social media algorithms have negative effects on mental health,
particularly for younger users, raises the question of whether AI assistants
optimised by the same engagement metrics will exacerbate these effects.
Meta’s internal research practices around user wellbeing have been the
subject of significant controversy, and the application of similar research
and optimisation approaches to AI assistant behaviour warrants independent
scrutiny that the company’s commercial interests give it limited incentive to
invite. For related analysis of AI company strategies, see our coverage of
the
costs of AI model training and LLMs
in everyday use.
The regulatory scrutiny facing Meta’s AI operations is
intensifying across multiple jurisdictions simultaneously. The EU’s AI Act
imposes specific obligations on providers of general-purpose AI models with
significant capabilities, including transparency requirements and safety
assessments that apply to the Llama model family. The UK’s Competition and
Markets Authority has initiated a review of foundation model markets that
includes assessment of Meta’s position. The US Federal Trade Commission has
ongoing investigations into Meta’s data practices that encompass its AI
development activities. The convergence of AI regulation across major markets
is creating compliance obligations that Meta, as a global platform company, must
manage across dozens of regulatory jurisdictions with different requirements.
How Meta navigates this regulatory environment, and whether its open-weight
strategy creates regulatory arbitrage opportunities that closed model
providers cannot access, will significantly shape the competitive dynamics of
the AI industry over the next several years. FTC disclosures from
ongoing proceedings provide some visibility into the regulatory concerns
about Meta’s data practices.
The AI assistant competition between
Meta, Google, Apple, and Microsoft is entering its most consequential phase,
and the outcome will shape how billions of people access information and make
decisions for years to come.
About the Author
Stuart Kerr is a technology correspondent at LiveAIWire, covering
artificial intelligence, digital innovation, and the social impact of emerging
technologies. Follow LiveAIWire for daily analysis at liveaiwire.com.