In brief
Meta is one of the rare major players to have made its language models widely accessible, under the Llama banner. This choice has fueled a global ecosystem of local deployments, fine-tuning, and academic research that closed laboratories have not been able to match in terms of reach. But this openness has precise limits, and Meta’s trajectory since 2025 shows a company in the process of recalibrating its relationship to open source under pressure from competition and infrastructure costs.
In short: think of Meta as a software vendor that distributes the “community” version of its product (Llama in open weights) for free while keeping the engine of future versions and the strategic direction in-house. The massive distribution creates ecosystem dependence — every company that builds on Llama integrates Meta’s technical choices — without giving up control. The economic model is not selling models, it is becoming the standard infrastructure on which others build.
Identity card
| Attribute | Value |
|---|---|
| FAIR founded | 2013 |
| Headquarters | Menlo Park, California |
| Parent company | Meta Platforms, Inc. |
| Key figures | Mark Zuckerberg (CEO), Alexandr Wang (Chief AI Officer since 2025) |
| Status | Publicly listed company (NASDAQ: META), integrated AI division |
History
FAIR, 2013: the lab as a signal
In 2013, Facebook created FAIR — Facebook AI Research — under the direction of Yann LeCun, a French researcher specializing in convolutional neural networks. The ambition was explicit: to rival the best university labs in the world, while remaining committed to open research. FAIR published prolifically, collaborated with universities, and built academic credibility that few technology companies possess.
LeCun led FAIR for five years, then held the position of Chief AI Scientist at Meta until November 2025 — twelve years in total within the organization. His departure marks a symbolic break: he left Meta to found AMI Labs (Advanced Machine Intelligence Labs), raising $1.03 billion in March 2026 to develop an alternative to autoregressive LLMs based on world models and the JEPA architecture.
Rebranding and generative pivot
In 2021, Facebook became Meta Platforms. FAIR was renamed Meta AI Research. The pivot toward the metaverse dominated the media space, but behind the scenes, teams were preparing what would become the Llama family.
In February 2023, Llama 1 was released — initially with restricted access for researchers. The weights leaked to BitTorrent a few weeks later. This episode, unintentional, prefigured the dynamic that would take hold: once weights are available, their distribution escapes Meta’s control.
2025 reorganization
Following results deemed disappointing for Llama 4, Zuckerberg restructured the AI organization. Meta invested $14.3 billion in Scale AI and recruited its CEO, Alexandr Wang, as Chief AI Officer. A new division was created: Meta Superintelligence Labs, with a mandate explicitly oriented toward commercial competitiveness rather than fundamental research.
In short: the 2025 reorganization marks three pivot points. First, $14.3 billion invested in Scale AI — the equivalent of an average European unicorn acquisition, just to put hands on a team and a CEO. Second, the symbolic departure of Yann LeCun, twelve years in the house, founder of FAIR — open fundamental research culture giving way to a product logic. Third, the internal rebranding (Superintelligence Labs) signals an explicit alignment on the AGI race, where Meta had always carefully avoided this vocabulary.
Models and products
The Llama family
The Llama family represents the public backbone of Meta’s AI strategy.
| Version | Date | Key points |
|---|---|---|
| Llama 1 | Feb. 2023 | Restricted access, weights leaked |
| Llama 2 | Jul. 2023 | Commercial use allowed, Microsoft partnership |
| Llama 3 | Apr. 2024 | 8B and 70B parameters |
| Llama 3.1 | Jul. 2024 | 405B parameters, 128K-token context |
| Llama 3.2 | Sep. 2024 | Vision models + 1B/3B multilingual models |
| Llama 4 | Apr. 2025 | MoE architecture, natively multimodal |
| Muse Spark | Apr. 2026 | First Superintelligence Labs model — closed weights |
| Muse Image | Jul. 2026 | Image generation |
| Muse Glimmer | Aug. 2026 | 30B dense distilled from Muse Spark, Apache 2.0, local agentic use |
Llama 4 introduces a Mixture-of-Experts architecture. The Scout model features 17B active parameters out of 109B total, with a context window of 10 million tokens. The Maverick model (17B active, 400B total) positions itself at the level of GPT-4o on several benchmarks. The Behemoth model — 288B active parameters, 2,000 billion total — had not yet been released at the time of writing.
SAM and computer vision
Meta also publishes influential vision models. SAM 3 (November 2025) is a unified model for detection, segmentation, and tracking via text, visual, or exemplar prompts. SAM Audio (December 2025) extends this paradigm to sound: extraction of audio sources from complex mixtures. These models are published with code, weights, and datasets.
Infrastructure
Meta does not merely publish models: the company is building massive infrastructure. 2025 CapEx: $70 to $72 billion. Estimated 2026 CapEx: $115 to $135 billion. A 2 GW datacenter with more than one million Nvidia GPUs is under construction. In January 2026, Zuckerberg announced that Meta is becoming an AI infrastructure provider (“Meta Compute”), with a plan to invest $600 billion in American AI infrastructure by 2028.
In short: to put the numbers in perspective. $70 billion of 2025 CapEx = about 30% of NASA’s annual budget, or the equivalent of the United Kingdom’s annual military budget. $600 billion by 2028 = the current market value of Tesla. 1 million Nvidia GPUs = roughly $30 billion in hardware alone. This scale places Meta in a category of its own: only Google, Microsoft and Amazon invest at this scale in AI infrastructure, and Meta does not have the cloud advantage of the other three.
Positioning
The open weights strategy: four motives, one effect
Meta structures its strategy around four arguments. First, security: open models benefit from community scrutiny. Second, standardization: when companies build on Llama and PyTorch, they naturally integrate Meta’s innovations. Third, talent: open source attracts researchers. Fourth, distribution: while OpenAI and Anthropic control their closed ecosystems, Meta influences the ecosystem of all other organizations.
The concrete effect is real. Llama has become the reference for local deployments, enterprise fine-tuning, and academic research on open models. Meta has captured an infrastructure position without building a commercial model-access platform.
Open weights ≠ open source
The distinction is fundamental. Llama models are not open source in the sense of the Open Source Initiative. Meta uses proprietary licenses (“Meta Llama Community License”) that carry significant restrictions: conditional commercial use, exclusion of companies exceeding 700 million monthly active users, restrictions on naming derivative models. Experts such as Nathan Lambert have documented the gap between Meta’s “open source” discourse and the reality of the licenses.
The limits of openness
Since mid-2025, signals indicate growing tension between the open source strategy and competitive requirements. TechCrunch headlined in July 2025: “Meta built its AI reputation on openness — that may be changing.” The creation of Meta Superintelligence Labs and the recruitment of Alexandr Wang mark a shift toward commercial performance at the expense of open fundamental research. Moreover, publishing weights does not guarantee transparency about training data, evaluation processes, or alignment choices — areas of opacity that the scientific community regularly highlights.
Key takeaways
- Meta occupies a singular position in the AI ecosystem: third global actor by resources, first actor by the reach of open model distribution.
- This position has been built methodically since 2013, with FAIR as the academic foundation and Llama as the distribution vehicle.
- What Meta publishes is open weights, not open source: Llama licenses carry significant usage restrictions.
- The departure of Yann LeCun and the creation of Meta Superintelligence Labs illustrate a drift toward commercial competitiveness at the expense of fundamental research.
- The scale of infrastructure investment — $70 billion in 2025, $115 to $135 billion projected in 2026 — confirms that Meta is playing a long game.