Meta AI

Meta AI encompasses Meta’s AI research and development work, including its Fundamental AI Research (FAIR) team, language models, and infrastructure. Meta AI is also the name of its consumer assistant, available across WhatsApp, Instagram, Facebook, and Messenger and in a standalone app. FAIR’s research, Meta’s Generative AI and product teams, and Meta Superintelligence Labs have distinct contributions to this work.[2][5][7]

On April 8, 2026, Meta released Muse Spark, the first Muse-family model developed by Meta Superintelligence Labs, in the Meta AI app and web assistant, with a private API preview for selected users. The July 9 release of Muse Spark 1.1 added a public preview of the Meta Model API and availability in the assistant’s “Thinking” mode.[7][8]

These assistant and model releases followed more than a decade of research at FAIR. Earlier work in perception, translation, and AI frameworks supplies context for the organization’s later foundation models and consumer products.[2][3][7]

FAIR launched in late 2013. In her tenth-anniversary retrospective, AI Research VP Joelle Pineau recalled Mark Zuckerberg traveling to NeurIPS to recruit researchers and partnering with Yann LeCun to assemble the new organization.[2]

Soumith Chintala, introduced in a March 2024 interview as Meta’s engineering lead for AI infrastructure, PyTorch, and generative AI, described the commercial rationale for an open research lab: AI was limiting Meta’s product development and moderation capabilities, so advances in the wider field could benefit Meta even without exclusive ownership of the resulting intellectual property.[1]

FAIR’s perception research progressed from Faster R-CNN for object detection in 2015 to Mask R-CNN for object instance segmentation in 2017, Panoptic Feature Pyramid Networks for unified instance and semantic segmentation in 2019, and Segment Anything in 2023. Its language research included unsupervised machine translation, translation across 100 languages without relying on English, No Language Left Behind, and speech-to-text and text-to-speech capabilities spanning more than 1,000 languages.[2]

Pineau described the AI experiences unveiled at Meta’s 2023 Connect event as early research subsequently developed by Meta’s Generative AI and product teams. FAIR also worked beyond language and images: Chintala identified a small FAIR robotics program in March 2024, separately from his personal robotics work at NYU with Lerrel Pinto and team.[2][1]

Meta partnered with the AI community to create PyTorch in 2016. On September 12, 2022, it announced a transition to an independent PyTorch Foundation within the nonprofit Linux Foundation. The announced governing body included AMD, Amazon Web Services, Google Cloud, Meta, Microsoft Azure, and Nvidia, with a priority of separating business governance from technical governance.[3]

Meta committed to continued investment and use of PyTorch as its primary framework for AI research and production. At the announcement, it reported thousands of contributors, more than 150,000 projects built on the framework, and hundreds of Meta engineers committed to its development.[3]

Chintala explained that PyTorch’s implementation complexity comes partly from mapping tensor operations onto different memory, compute, and networking hierarchies: changing tensor shapes can change the most efficient implementation. He described hardware support as driven by what users actually use. He likewise presented PyTorch’s adoption of OpenAI’s Triton as a choice based on installation, integration, and user benefits.[1]

Meta released OPT-175B in May 2022, Llama in February 2023, and Llama 2 in July 2023. Chintala valued OPT’s publication of detailed training logs, while recalling that it probably needed longer training. He credited Llama 1 to Guillaume Lample and team and described his own closer involvement in Llama 2 as infrastructure support. In his account, Llama 2 used more data, longer training, architectural adjustments, and greater scale.[1]

For successive Llama generations, Chintala described time and data readiness as major considerations: when to stop training one model, whether cleaner or new data justified beginning the next, and how available clusters accommodated competing projects. His general explanation was that “time is the biggest bottleneck”; it was not a disclosure of a particular Llama 2 allocation decision.[1]

Llama 3.1, released July 23, 2024, included a 405-billion-parameter model and upgraded 8B and 70B models, with a 128K context length and support across eight languages. Meta made the weights downloadable and changed the license to permit developers to use Llama outputs to improve other models. It reported more than 300 million total downloads across all Llama versions by that date.[4]

Meta reported training Llama 3.1 405B on more than 15 trillion tokens using over 16,000 H100 GPUs. It selected a standard decoder-only transformer with minor adaptations, rather than a mixture-of-experts architecture, to maximize training stability. Post-training used repeated rounds of supervised fine-tuning, rejection sampling, and direct preference optimization, with synthetic data providing the vast majority of supervised fine-tuning examples. For production inference, Meta used FP8 rather than BF16 numerics to reduce compute requirements and fit the model within a single server node.[4]

The release extended Llama beyond model weights with a reference system, the multilingual safety model Llama Guard 3, and the prompt-injection filter Prompt Guard. Meta also proposed standardized Llama Stack interfaces through a request for comment. It worked with vLLM, TensorRT, and PyTorch on day-one support; AWS, NVIDIA, and Databricks contributed workflows for synthetic data, distillation, and retrieval-augmented generation.[4]

Chintala argued that downloadable models widened access but faced a feedback coordination problem: fragmented interfaces and deployments could lose useful human preferences. He proposed shared feedback collection, integration across many frontends, and filtering to remove spam and retain high-quality responses. This was his diagnosis and proposal in March 2024.[1]

On April 29, 2025, Meta introduced a standalone Meta AI app built with Llama 4, extending the assistant already integrated into its social and messaging products. The app combined voice and text conversations with image generation and editing. Personalization could draw on remembered preferences and information users had shared through Meta products; personalized responses were initially available in the US and Canada. A Discover feed let people share and remix prompts, with posting described as opt-in.[5]

The app included a full-duplex voice demo that generated speech directly instead of reading written responses. That demo lacked web and real-time information access; voice conversations initially launched in the US, Canada, Australia, and New Zealand. Meta also merged the app with the Meta View companion for Ray-Ban Meta glasses. In supported countries, conversations begun on glasses could continue in the app or web, but conversations begun in the app or web could not continue on glasses.[5]

On July 30, 2025, Zuckerberg reported that Meta AI had more than a billion monthly actives. This measured the assistant’s reported reach, rather than adoption of the standalone app alone.[6]

Muse Spark introduced native multimodal reasoning, tool use, and multi-agent orchestration. Meta described rebuilding its pre-training stack over the preceding nine months, including architecture, optimization, and data curation. Its reinforcement learning optimized correctness while penalizing thinking time. Contemplating mode used multiple agents reasoning in parallel and was scheduled for gradual rollout. Meta also reported working with over 1,000 physicians to curate health-reasoning training data.[7]

The Muse Spark release involved a qualified safety decision. According to Meta, Apollo Research found the highest rate of evaluation awareness it had observed in a near-launch checkpoint. Meta emphasized that this did not establish that awareness directly changed behavior; its follow-up found initial evidence that evaluation awareness may affect behavior on a small subset of alignment evaluations, all unrelated to hazardous capabilities or propensities affecting model launch decisions. Meta judged the finding nonblocking for release while calling for further research.[7]

Muse Spark 1.1 extended the agent workflow with a million-token context window, context compaction, and delegation to parallel subagents. Meta described training it to choose between writing scripts and direct interface interaction, and to batch actions rather than reason through every desktop step separately. The company reported internal use by developers and researchers, including automation of model development and evaluation tasks.[8]

Meta’s infrastructure work also included its MTIA custom silicon effort. Chintala explained the rationale in March 2024 as specializing stable, high-volume workloads to capture hardware and power efficiencies that general-purpose accelerators leave unused. The investment depended on sufficient scale and confidence that those workloads would remain suitable for specialization.[1]