🇺🇸 United States · Meta AI — Facebook’s AI Research and Products
Status: 🟩 COMPLETE 🟦 LIVING Section: 10 — AI and LLMs
| Vendor | Meta Platforms, Inc. (formerly Facebook) |
| Country/origin | 🇺🇸 United States (Menlo Park, California) |
| Recommended for AUS? | ⚠️ Use with awareness — US company; significant social media data integration; see privacy section |
| Privacy summary | Meta’s standard privacy practices apply; social media data used for training; strong open-weights commitment; no AI-specific data isolation by default |
| Free tier | Yes — Meta AI assistant is free; Llama models are open-weights (free to download) |
| Paid tiers | No consumer AI subscription; API access via cloud providers |
| First released | Meta AI (assistant) launched September 2023; Llama 1 released February 2023 |
| Last reviewed | June 2026 |
| Official site | https://ai.meta.com |
What it is
Meta is the company behind Facebook, Instagram, WhatsApp, and Ray-Ban smart glasses — and one of the world’s most significant AI research organisations. Meta’s AI activities are organised under two primary groups:
-
FAIR (Fundamental AI Research): Meta’s foundational AI research lab, founded 2013. Produces research that is typically published openly and has contributed foundational work to the field — the original transformer implementations, image recognition benchmarks, and much more. FAIR is one of the most respected AI research groups in the world.
-
GenAI (Generative AI): Meta’s product-focused AI team, responsible for building the AI features that appear in Facebook, Instagram, WhatsApp, and the Meta AI assistant.
Key products:
- Meta AI assistant: Integrated into Facebook, Instagram, WhatsApp, and Messenger. Powered by Llama models. Free, no account required beyond your existing Meta account.
- Llama (open-weights models): Meta’s most significant contribution to the AI ecosystem. Released as open-weights (downloadable, runnable locally, modifiable). See llama for full detail.
- ImageBind: A research model that understands and connects multiple types of data (images, audio, text, depth, video) simultaneously
- SAM (Segment Anything Model): Open-source model for identifying and cutting out objects in images — widely used in image editing software
- AudioCraft/MusicGen: Open-source music and audio generation models
- Ray-Ban Meta smart glasses AI: AI assistant built into glasses hardware
Meta’s unique position in the AI landscape
Meta occupies a very different position from other AI companies:
-
Open-weights commitment: Meta releases its AI models freely under open licences (Llama community licence). This is a deliberate strategic choice — partly to commoditise the AI model layer so competitors (OpenAI, Google) can’t charge premium prices for models, partly to build developer goodwill and ecosystem.
-
Scale of deployment: Meta’s AI reaches billions of people through Facebook, Instagram, and WhatsApp — making it one of the most widely-deployed AI systems in the world even if it’s not the most sophisticated.
-
Social media data advantage: Meta has access to enormous amounts of human-generated content from its platforms — conversations, images, videos. This is both a training data advantage and a privacy concern.
-
Hardware investment: Meta is one of the world’s largest buyers of NVIDIA GPUs and is developing its own custom AI chip (MTIA — Meta Training and Inference Accelerator).
The Llama model family
| Model | Release | Notable |
|---|---|---|
| Llama 1 | Feb 2023 | First major open-weights LLM release; 7B–65B |
| Llama 2 | Jul 2023 | More permissive licence; commercial use allowed |
| Llama 3 | Apr 2024 | Dramatic quality improvement; competitive with proprietary models |
| Llama 3.1 405B | Jul 2024 | First open-weights model competitive with GPT-4 class |
| Llama 3.2 | Sep 2024 | Multimodal (vision); small models for on-device |
| Llama 3.3 | Dec 2024 | Further improvements to 70B and 405B |
| Llama 4 | 2025 | Mixture-of-experts architecture; next-level capability |
FAIR research highlights
FAIR has produced landmark research including:
- OPT (Open Pre-trained Transformers, 2022): One of the first large language models released openly
- Segment Anything (SAM, 2023): Revolutionised image segmentation
- AudioCraft / MusicGen (2023): High-quality music generation models
- LLM for biology (ESMFold): Protein structure prediction
- No Language Left Behind (NLLB, 2022): Translation model for 200+ languages including low-resource languages
Privacy and data considerations
Meta’s AI activities are intertwined with its social media businesses, which raises specific considerations:
- Training data: Meta has stated it trains AI models using public posts from Facebook and Instagram. In some jurisdictions, users had to opt out of this (EU regulators required this; Australian users should check current terms).
- Meta AI in conversations: If you use Meta AI in WhatsApp or Messenger, those conversations may be used to improve Meta’s AI. Check privacy settings.
- Data integration: Unlike standalone AI companies, Meta’s AI is part of an ecosystem designed to build advertising profiles. The two goals (AI and advertising) create inherent tensions.
- Ray-Ban smart glasses: AI cameras on glasses raise specific public-space privacy concerns around bystander footage.
Australian context: The Australian Privacy Act 1988 applies to Meta’s operations in Australia. The OAIC (Office of the Australian Information Commissioner) has previously investigated Meta. Check Meta’s Australian privacy settings and opt-outs.
How Meta AI (the assistant) compares to alternatives
| Assistant | Quality | Integration | Privacy |
|---|---|---|---|
| Meta AI | Good | Facebook/Instagram/WhatsApp | ⚠️ Social media data |
| ChatGPT | Excellent | Standalone; broad plugins | Moderate |
| Google Gemini | Excellent | Google ecosystem | Moderate |
| Claude | Excellent | Standalone; enterprise options | Good |
| Grok | Very good | X/Twitter | ⚠️ Musk ownership |
Gotchas
- Meta AI is the product; Llama is the model. Many people confuse the two. Meta AI is the consumer assistant in Facebook/Instagram/WhatsApp. Llama is the underlying model — available separately, usable in completely different applications.
- Llama licences have restrictions. Llama models are “open-weights” but not fully “open-source” by the traditional definition. The community licence restricts commercial use above certain scale (100M+ monthly users require a separate licence from Meta).
- Meta AI quality varies by platform. The experience in WhatsApp differs from Facebook, and the model may differ from what’s available via API.
- Social media data concerns are real. If you’re using Meta AI for sensitive business discussions via Messenger or WhatsApp, be aware that Meta’s data practices apply.
- Australian opt-out of AI training: Check Meta’s Privacy Settings → Privacy Policy → AI information section to understand your current opt-in/opt-out status in Australia.
See also
- llama — the open-weights model family (full entry)
- meta-ai — the consumer assistant (full entry)
- openai-company — comparison company
- anthropic — comparison company
- vendors-western-recommended — note the ⚠️ on Meta
Sources
- Meta AI Research (FAIR): ai.meta.com/research
- Llama model release announcements and papers (2023–2026)
- Meta AI product announcements (2023–2026)
- OAIC (Office of the Australian Information Commissioner) — Meta investigation outcomes
- Privacy International analysis of Meta AI (2024)
- TechCrunch, The Verge comprehensive Meta AI coverage (2023–2026)