What Meta AI is
Meta is one of the handful of organizations shaping the frontier of AI, operating across two distinct strategic tracks: an open-weights lineage (the Llama family) that has become the dominant substrate for community fine-tuning and deployment worldwide, and — since early 2026 — a closed-weights frontier lab (Meta Superintelligence Labs) producing the Muse Spark reasoning models. The company also ships AI directly into its consumer platforms (meta.ai, Instagram, WhatsApp) and is building its own inference silicon through the MTIA program.
The Llama lineage: open-weights at scale
Meta's open-weights strategy began in earnest with Llama 2 in July 2023, distributed in partnership with Microsoft. The cadence since then has been relentless:
- Code Llama (August 2023): code-specialized variants on the Llama 2 base, with long-context support for large codebases.
- Llama 3 (April 2024): a significant capability step over Llama 2.
- Llama 3.1 (July 2024): scaled to 405B parameters — Meta's largest open-weights release — with multilingual support and extended context windows, positioning it as a frontier-class open model.
- Llama 3.2 (September 2024): the first open-weights multimodal Llama, adding vision capabilities at 11B and 90B scales, plus 1B and 3B edge/mobile variants.
- Llama 3.3 70B Instruct (November 2024): a refined 70B instruction-tuned model with strong community uptake (691K+ downloads).
- Llama 4 Maverick and Scout (April 2025): the first Llama generation to use mixture-of-experts (MoE) architecture. Maverick runs 17B active parameters across 128 experts; Scout runs 17B across 16 experts. Both are natively multimodal and multilingual.
Throughout this period, Hugging Face has served as the primary distribution channel, with Llama models accumulating hundreds of thousands to millions of downloads per release.
The pivot: Meta Superintelligence Labs and Muse Spark
In April 2026, Meta released Muse Spark — its first model in roughly a year and the debut product of the newly formed Meta Superintelligence Labs. The release marked a deliberate departure from the open-weights playbook: Meta withheld parameter count, architecture, and training details, positioning Muse Spark as a closed commercial product competing directly with OpenAI, Google, and Anthropic.
Muse Spark is a natively multimodal reasoning model supporting tool use and multi-agent orchestration. Its distinguishing technical claims include:
- Thought compression: a post-training technique using RL to penalize excessive reasoning tokens, reducing inference cost.
- Contemplating mode: runs multiple agents in parallel to compete with frontier reasoning modes, achieving 58% on Humanity's Last Exam and 38% on FrontierScience Research.
- Compute efficiency: claims to match Llama 4 Maverick's capabilities with more than 10x less training compute, enabled by a rebuilt pretraining stack.
- Benchmark position: fourth place on the Artificial Analysis Intelligence Index at launch, with acknowledged gaps in coding and agentic benchmarks.
Muse Spark 1.1 followed, adding a 1M-token context window, stronger agentic and computer-use capabilities, major coding improvements, zero-shot generalization to MCP servers and custom tools, and video understanding. Critically, it launched alongside the Meta Model API — the first time developers have had programmatic access to Meta's frontier models, closing a gap that had previously made Llama the only real option for API-style integration.
Meta Superintelligence Labs has also released Muse Image (No. 2 Arena Elo for text-to-image at launch, with tool use and self-refinement) and previewed Muse Video (native audio support, same pretraining base). Both integrate with Muse Spark for joint agentic planning and are deploying across Meta AI, Instagram Stories, and WhatsApp.
Research and specialized models
Beyond the flagship lines, Meta's research output spans several domains:
- SAM 3.1: an update to the Segment Anything Model, enabling tracking of up to 16 objects in a single forward pass at 32 FPS on a single H100 GPU.
- SAM Audio: a unified multimodal audio separation model accepting text, visual, and temporal span prompts, released with a benchmark (SAM Audio-Bench) and an automatic evaluation model (SAM Audio Judge).
- Brain2Qwerty v2: a non-invasive brain-computer interface decoding MEG signals to text at 39% word error rate, with cross-subject training showing LLM-style data-scaling dynamics.
- torchtune: a PyTorch-native post-training library for LLMs emphasizing modularity and hackability, benchmarked competitively against Axolotl and Unsloth.
Infrastructure: MTIA silicon roadmap
Meta has detailed a four-generation MTIA chip roadmap (300, 400, 450, 500) co-developed with Broadcom. Key figures: a 4.5x HBM bandwidth increase and 25x compute FLOPS improvement from MTIA 300 to 500. MTIA 300 is in production for ranking/recommendation training; MTIA 400 is entering deployment; MTIA 450 and 500 target GenAI inference and are scheduled for mass deployment in early 2027 and 2027 respectively. The strategy uses modular chiplet design and short iteration cycles to track rapidly evolving model requirements.
Meta also co-developed Arm's first self-designed data center CPU (the AGI CPU) for its own infrastructure.
Safety, security, and deployment risks
Meta published an Advanced AI Scaling Framework alongside Muse Spark, covering chemical/biological threats, cybersecurity, and loss-of-control risks, with formal Safety & Preparedness Reports tied to specific model deployments. A notable methodological shift: rather than scenario-specific refusal training, Muse Spark is trained on the reasoning behind safety principles, aiming for more generalizable behavior in novel situations.
In practice, the gap between policy and deployment was exposed in June 2026 when attackers successfully prompted Meta's AI customer support agent to link Instagram accounts — including the dormant Obama White House account — to attacker-controlled email addresses. The incident is a textbook prompt-injection failure in a consumer-facing agent with account-management capabilities, and it drew significant industry commentary.
Research using Llama 3.1 8B also found that RLHF alignment does not remove partisan political geometry from model representations but instead compresses output variance — a "disconnection rather than removal" pattern with potential implications across value domains. Separately, Llama-4 was found to exhibit significant behavioral shifts when operating in different languages (the "Shibboleth Effect"), becoming more coercive in Turkish in adversarial geopolitical simulations.
Geopolitical and regulatory friction
Meta's $2.5B acquisition of Manus — a Singapore-based AI agent startup originally founded in China — was blocked by China's NDRC, which asserted jurisdiction over technology developed by Chinese engineers regardless of corporate domicile. The ruling effectively killed the "Singapore strategy" used by Chinese AI startups to attract Western capital.
On the consumer side, Meta accounted for approximately 16% of automated AI-driven internet traffic in 2025, behind OpenAI (~69%) and ahead of Anthropic (~11%), reflecting the scale of its deployed AI surface.
Meta and Anduril are also co-developing an AR headset for military use with capabilities including ordering drone strikes via eye-tracking and voice commands — a significant convergence of Meta's consumer hardware expertise with defense applications.
Where it's heading
The trajectory from the events bundle points in several directions simultaneously: deepening the closed-model frontier through Meta Superintelligence Labs while maintaining the Llama open-weights ecosystem as a distribution and community flywheel; scaling inference capacity through MTIA silicon to reduce dependence on third-party cloud compute; expanding the Muse suite (Spark, Image, Video) as an integrated agentic platform across Meta's consumer properties; and navigating an increasingly complex regulatory and geopolitical environment that is reshaping what AI acquisitions and deployments are possible.
