Meta AI: Llama & Muse Spark Models Explained (2026)
Meta AI, founded 2013, builds Llama (1B+ downloads) and Muse Spark (262K context, #5 Arena Elo). Spending $145B on AI in 2026. Free and API access options.
Meta AI is Meta Platforms' AI division (founded 2013), creator of the Llama open-weight family with 1 billion+ downloads and Muse Spark, released April 8, 2026, with a 262K-token context window and 1491 Elo on Chatbot Arena. Meta Platforms (NASDAQ: META) is investing up to $145 billion in AI infrastructure in 2026, making it one of the three largest AI spenders globally alongside Microsoft and Google.
Meta AI, founded in 2013 as Facebook AI Research (FAIR), is Meta Platforms' artificial intelligence division and the maker of the Llama open-weight model family (1B+ downloads) and Muse Spark, its first closed proprietary frontier model released April 8, 2026. Muse Spark reached #5 on Chatbot Arena with 1491 Elo and features a 262K-token context window. Meta is spending up to $145 billion on AI infrastructure in 2026.
Founded: 2013 · HQ: Menlo Park, California, USA · Team: 10,000+ · CEO: Mark Zuckerberg (CEO, Meta Platforms); Alexandr Wang (Chief AI Officer, MSL) · Funding: Public (NASDAQ: META); $14.3B invested in Scale AI/MSL; $145B 2026 AI capex · Valuation: $1.4T+ (Meta Platforms market cap, mid-2026)
About Meta AI
Meta AI is the artificial intelligence division of Meta Platforms, Inc., the company formerly known as Facebook. It was established in 2013 as the Facebook Artificial Intelligence Research (FAIR) lab, originally led by Yann LeCun, who remains one of the most cited researchers in deep learning. Over the following decade, FAIR produced influential work in computer vision, speech recognition, and natural language processing before transitioning into a product-centric division that ships frontier models at consumer scale. Meta's AI strategy has two distinct tracks. The first is the Llama open-weight family, which Meta began releasing in February 2023. Llama models are available for download under a commercial use license and have accumulated over one billion downloads as of early 2026, making Meta the dominant force in the open-weights ecosystem. The Llama 4 generation, released in April 2025, introduced mixture-of-experts architecture and native multimodal inputs, competing directly with Anthropic, Google, and OpenAI on benchmarks while remaining free to download for most use cases. The second track is the Muse family, launched on April 8, 2026, by Meta Superintelligence Labs (MSL). Muse Spark, the first model in the family, is the company's first fully closed proprietary model. It is natively multimodal, handling text, images, and voice in a unified architecture, and introduces Contemplating Mode, which orchestrates multiple reasoning agents in parallel. This departure from the open-weights strategy reflects Meta's belief that the frontier of AI capability now requires keeping model weights proprietary to maintain competitive advantage. Meta Superintelligence Labs was announced by CEO Mark Zuckerberg in June 2025, following Meta's $14.3 billion investment in Scale AI and the appointment of Scale AI's founder, Alexandr Wang, as Meta's first Chief AI Officer. Wang co-leads MSL with Nat Friedman, former head of GitHub, and Shengjia Zhao serves as chief scientist. The lab is internally structured into four groups: the TBD Lab for large language models, FAIR for fundamental research, a products and applied research division, and an infrastructure unit led by Aparna Ramani. Meta's infrastructure investment for 2026 is the largest in the company's history. Full-year capital expenditure guidance was raised to between $125 billion and $145 billion, covering data center construction, custom AI silicon, and energy procurement. This level of spending puts Meta in the same tier as Microsoft and Google as hyperscalers betting the company on AI dominance. Consumer-facing AI products include the Meta AI assistant, integrated into Facebook, Instagram, WhatsApp, and Messenger, with over 700 million monthly active users by early 2026. Meta AI is also embedded in Ray-Ban AI smart glasses, which sold 7 million units in 2025, making wearable AI an unexpected growth product. The assistant is powered by Muse Spark as of April 2026, replacing the previous Llama-based backend. On the enterprise and developer side, Meta's API access for Muse Spark remained in private preview through mid-2026, with no public pricing announced. Llama models, by contrast, are broadly available through Together AI, Fireworks AI, AWS Bedrock, and Google Vertex AI. Meta's open-weight strategy has created a large ecosystem of fine-tuned variants, quantized builds, and specialized deployments that no proprietary API could replicate at the same cost. Meta Platforms is a public company listed on NASDAQ under the ticker META, with a market capitalization exceeding $1.4 trillion as of mid-2026. The company generated $47.5 billion in net income in 2025, funding AI investments without external fundraising. Unlike OpenAI, Anthropic, or Mistral, Meta AI does not raise venture capital for its AI division. This financial structure gives Meta freedom to pursue long research timelines and absorb infrastructure costs that would be prohibitive for startup competitors.
Mission
Build personal superintelligence safely and make it available to everyone.
Products
- Muse Spark (model): https://ai.meta.com/blog/introducing-muse-spark-msl/
- Llama 4 (model): https://ai.meta.com/blog/llama4/
- Meta AI Assistant (product): https://ai.meta.com/
- Ray-Ban AI Glasses (hardware): https://www.ray-ban.com/usa/meta-smart-glasses
Compliance
GDPR, CCPA
Links
Website · GitHub · Twitter · LinkedIn · Blog · Docs
Frequently Asked Questions
What is Meta AI and who leads it?
Meta AI is the artificial intelligence division of Meta Platforms, Inc., originally founded in 2013 as the Facebook Artificial Intelligence Research (FAIR) lab under Yann LeCun. Today it operates under two units: FAIR (fundamental research) and Meta Superintelligence Labs (MSL), which was announced in June 2025. MSL is led by Alexandr Wang (Chief AI Officer, formerly founder of Scale AI) and Nat Friedman (former GitHub CEO) as co-leads, with Shengjia Zhao (ex-OpenAI) as chief scientist. CEO Mark Zuckerberg remains the overall decision-maker. As of 2026, Meta AI employs over 10,000 researchers and engineers across offices in Menlo Park, London, New York, Paris, Seattle, Tel Aviv, and Montreal.
What models does Meta AI make?
Meta AI maintains two parallel model lines. The first is Llama, an open-weight family that began with Llama 1 in February 2023 and grew through Llama 2, Llama 3, and Llama 4 (released April 2025). Llama 4 uses a mixture-of-experts architecture with native multimodal support and is available for download under a commercial license, accumulating over 1 billion downloads across versions. The second line is the Muse family, a new proprietary series launched by Meta Superintelligence Labs. Muse Spark, released April 8, 2026, is the first model in this family. It is closed-source, natively multimodal (text, image, voice), and reached #5 on Chatbot Arena with 1491 Elo. A public API for Muse Spark was still in private preview as of mid-2026.
How much is Meta spending on AI in 2026?
Meta raised its full-year 2026 capital expenditure guidance to between $125 billion and $145 billion, the largest single-year infrastructure investment in the company's history. This spending covers data center construction, custom AI silicon development, networking, and energy procurement. Meta funds this investment entirely from operating cash flow — the company generated $47.5 billion in net income in 2025 — without needing to raise venture capital or take on significant debt. This financial structure gives Meta the freedom to pursue multi-year research timelines that would be impractical for startup labs. For comparison, Microsoft committed roughly $80B and Google roughly $75B in AI-related capex for 2026.