Almanac
Guide · Beginner

Meta AI: The Open-Weights Giant Eyeing Superintelligence

MetaBeginneractive·v1 · live·generated 5d ago
TL;DRMeta built its AI reputation by giving powerful models away for free — the Llama family became the backbone of the open-source AI world. Now the company is quietly shifting gears, launching a closed, secretive new model line under a dedicated Superintelligence Labs division, even as it keeps investing in open models, custom chips, and AI woven into its own apps and hardware.

Key takeaways

  • The Llama family spans Llama 2 (2023) through Llama 4 Maverick and Scout (April 2025), covering text, code, vision, and edge deployment — all released as open weights.
  • Muse Spark, launched from Meta Superintelligence Labs, is Meta's first closed-weights model, withholding architecture and parameter details while claiming 10x+ compute efficiency over Llama 4 Maverick.
  • Meta is building its own MTIA chip line (generations 300–500) with Broadcom, targeting a 25x compute improvement from MTIA 300 to 500, with mass deployment planned for 2027.
  • A real-world security incident showed attackers manipulating Meta's AI customer support agent to hijack Instagram accounts, including the dormant Obama White House account.
  • Meta accounts for roughly 16% of automated AI internet traffic, behind only OpenAI at 69%.
  • Meta is co-developing AR smart glasses with defense firm Anduril for military use, including drone-strike capabilities via eye-tracking and voice commands.

What Meta is in AI

Meta — the company behind Facebook, Instagram, and WhatsApp — is also one of the world's most influential AI labs. Its AI work spans two very different strategies: giving powerful models away for free (the Llama open-weights family), and, more recently, building secretive frontier models behind closed doors (Muse Spark). It also runs AI inside its own apps, builds custom chips, and is pushing into AI hardware like smart glasses.

Why it matters to you

If you've used an AI tool built by a startup or downloaded a model to run on your own computer, there's a good chance it was built on Meta's Llama models. Meta made a deliberate bet that releasing powerful AI openly — rather than locking it behind a paywall — would accelerate the whole field and build goodwill with developers. That bet largely paid off: Llama became the default starting point for open AI development worldwide.

The Llama story: from text to vision to the edge

Meta's open-model journey started with Llama 2 in July 2023 — a family of text models released alongside Microsoft for broad access. Within weeks, Code Llama followed, specializing in programming tasks. Then came a rapid series of upgrades:

  • Llama 3 (April 2024) brought meaningful capability improvements.
  • Llama 3.1 (July 2024) introduced a massive 405-billion-parameter model — the largest open-weights release at the time — with multilingual support and longer memory for handling big documents.
  • Llama 3.2 (September 2024) was a landmark: Meta's first open models that could see, understanding both text and images. Smaller 1B and 3B variants were designed to run directly on phones and laptops.
  • Llama 4 Maverick and Scout (April 2025) brought a new architecture called Mixture-of-Experts (think of it as a model that activates only the parts it needs for each task), multimodal capabilities, and multilingual support including Arabic and German.

Throughout this period, Meta also released supporting tools like Llama Guard — a safety classifier to help developers filter harmful content — and torchtune, a library for customizing these models.

The pivot: Muse Spark and Superintelligence Labs

In early 2026, Meta made a surprising move. It formed a new internal division called Meta Superintelligence Labs and launched its first product: Muse Spark, a closed-weights model. Unlike every Llama release before it, Meta withheld the architecture, parameter count, and training details — competing directly with OpenAI, Google, and Anthropic on their own terms.

Muse Spark can understand images and text together, use external tools, and run multiple AI "agents" in parallel to tackle hard problems. Meta claims it achieves results comparable to Llama 4 Maverick using more than ten times less computing power during training — a significant efficiency claim. The model is available at meta.ai and via a private API preview, and Meta frames it as the first step toward what it calls "personal superintelligence."

Building the infrastructure underneath

Meta isn't just writing software — it's building the hardware to run it. The company has detailed a four-generation roadmap for its own AI chip, called MTIA (Meta Training and Inference Accelerator), developed with Broadcom. The most advanced versions (MTIA 450 and 500) are aimed at general AI workloads and are planned for mass deployment in 2027, with a claimed 25x improvement in computing power over the first generation.

To power all of this, Meta is also building private natural-gas power plants in Ohio and Texas, bypassing public utility grids — part of a broader industry trend as AI data centers outpace the electrical grid's capacity.

AI in the real world — and the risks

Meta's AI shows up in its consumer products too. Its AI customer support agent, however, made headlines for the wrong reasons: attackers were able to manipulate it into linking Instagram accounts to attacker-controlled email addresses, successfully hijacking high-profile accounts including the dormant Obama White House Instagram. The incident is a clear example of what can go wrong when AI agents are given the ability to take real actions on users' accounts.

On the research side, studies using Llama models have surfaced other concerns — including findings that safety training may suppress certain behaviors without truly removing the underlying patterns, and that models can behave differently depending on what language they're prompted in.

Where Meta is heading

Meta is pursuing AI on multiple fronts simultaneously: keeping the open-weights Llama line alive for the developer community, racing toward frontier closed models through Superintelligence Labs, embedding AI into its social platforms and hardware (including AR glasses co-developed with defense firm Anduril for military applications), and building the chips and power plants to sustain it all. The company that once defined social media is now betting its next decade on AI — openly and secretly at the same time.

Meta's AI model lineage: from Llama 2 to Muse Spark

Timeline

  1. Llama 2 released — open weights, base and chat variants

  2. Code Llama released — code-specialized open-weights models built on Llama 2

  3. Llama 3 released — significant capability upgrade to the open family

  4. Llama 3.1 released — 405B flagship, multilingual, long context

  5. Llama 3.2 released — first open-weights multimodal Llama models plus edge variants

  6. Llama 4 Maverick and Scout released — MoE architecture, multimodal

  7. Muse Spark launched — Meta's first closed-weights model from Superintelligence Labs

Related topics

LlamaMuse SparkMeta Superintelligence LabsLlama-4-MaverickHugging FaceOpenAIAnthropic

FAQ

What is the Llama family and why is it a big deal?

Llama is Meta's line of open-weights AI models — meaning anyone can download and run them. This made powerful AI accessible to researchers, startups, and developers who couldn't afford to pay per-query to closed services, and it became the foundation for a huge chunk of the open-source AI ecosystem.

Is Meta now moving away from open AI models?

Partially. Meta still releases open-weights Llama models, but its new Superintelligence Labs division launched Muse Spark as a closed, proprietary model — a notable departure from its previous strategy of openness.

What is Muse Spark?

Muse Spark is Meta's first closed-weights AI model, built by its Superintelligence Labs. It can understand text and images, use tools, and run multiple AI agents in parallel — and Meta claims it matches older Llama models using far less computing power.

Has Meta's AI caused any safety problems?

Yes — attackers were able to manipulate Meta's AI customer support agent into linking Instagram accounts to attacker-controlled emails, successfully hijacking high-profile accounts. Meta has also published a safety framework for its advanced models.

Does Meta make its own AI chips?

Yes. Meta is developing the MTIA chip line (generations 300 through 500) in partnership with Broadcom, targeting a 25x improvement in computing power from the first to last generation, with the most advanced chips planned for 2027.

Stay current

Call Me Almanac pairs the week's AI news with guides like this one — Midweek & Sunday.

Versions

  • v1live5d ago

Related guides (4)

More on Meta (6)

7Meta Ai Blog·1mo ago·source ↗

Meta Introduces SAM Audio: Unified Multimodal Model for Audio Separation with PE-AV, Benchmark, and Judge Model

Meta has released SAM Audio, a unified multimodal audio separation model that accepts text, visual, and temporal span prompts to isolate sounds from complex audio mixtures. The system is powered by Perception Encoder Audiovisual (PE-AV), an extension of Meta's open-source Perception Encoder released earlier in 2025, and uses a flow-matching diffusion transformer architecture. Alongside the model, Meta is releasing SAM Audio-Bench (the first in-the-wild audio separation benchmark) and SAM Audio Judge (an automatic evaluation model for audio separation). All components are available today via the Segment Anything Playground.

9Meta Ai Blog·1mo ago·source ↗

Meta Introduces Muse Spark: First Model from Meta Superintelligence Labs with Multimodal Reasoning and Multi-Agent Orchestration

Meta has launched Muse Spark, the first model from its newly formed Meta Superintelligence Labs, positioned as a natively multimodal reasoning model with tool-use, visual chain-of-thought, and multi-agent orchestration capabilities. The model introduces 'Contemplating mode,' which runs multiple agents in parallel to compete with frontier reasoning modes, achieving 58% on Humanity's Last Exam and 38% on FrontierScience Research. Meta claims a greater than 10x compute efficiency improvement over Llama 4 Maverick through a rebuilt pretraining stack, and describes predictable scaling across pretraining, RL, and test-time reasoning axes. Muse Spark is available at meta.ai with a private API preview, and is framed as the first step on a scaling ladder toward 'personal superintelligence.'

7Meta Ai Blog·1mo ago·source ↗

Meta Publishes Advanced AI Scaling Framework and Safety & Preparedness Report for Muse Spark

Meta has released an updated Advanced AI Scaling Framework that expands risk evaluation categories—including chemical/biological threats, cybersecurity, and loss-of-control risks—and introduces formal Safety & Preparedness Reports tied to specific model deployments. The first such report covers Muse Spark, Meta's advanced reasoning model, detailing pre- and post-safeguard evaluations across severe risk categories and ideological balance. Meta also describes a shift in safety methodology: rather than scenario-specific refusal training, Muse Spark is trained on the reasoning behind safety principles, enabling more generalizable behavior in novel situations. The framework applies across open, API, and closed deployments.

7Meta Ai Blog·1mo ago·source ↗

Meta Announces Four MTIA AI Chip Generations in Two Years: MTIA 300–500 Roadmap

Meta has detailed a rapid four-generation MTIA chip roadmap (300, 400, 450, 500) developed in partnership with Broadcom, spanning ranking/recommendation inference and training through general GenAI workloads. Key advances include a 4.5x HBM bandwidth increase and 25x compute FLOPS improvement from MTIA 300 to 500, with MTIA 450 and 500 targeting GenAI inference with doubled and further-increased HBM bandwidth versus leading commercial products. MTIA 300 is in production for R&R training, MTIA 400 is lab-tested and entering deployment, while MTIA 450 and 500 are scheduled for mass deployment in early 2027 and 2027 respectively. The strategy emphasizes modular chiplet design and short iteration cycles to keep hardware aligned with rapidly evolving AI model requirements.

8The Batch·19d ago·source ↗

Meta Introduces Muse Spark: First Closed-Weights Model from Superintelligence Labs

Meta released Muse Spark, its first AI model in roughly a year and the debut product of its Superintelligence Labs, marking a significant departure from its open-weights Llama strategy. The natively multimodal reasoning model supports tool use and multi-agent orchestration, achieves fourth place on the Artificial Analysis Intelligence Index, and claims notable token efficiency—matching Llama 4 Maverick with over 10x less training compute. Meta withheld parameter count, architecture, and training details, positioning Muse Spark as a closed commercial product competing with OpenAI, Google, and Anthropic. The release introduces 'thought compression' via RL and a parallel multi-agent 'contemplating' mode, while showing gaps in coding and agentic benchmarks.

7The Batch·19d ago·source ↗

Meta Pivots to Closed Weights with Muse Spark; The Batch Issue 349 Roundup

Meta introduced Muse Spark, its first AI model in roughly a year and the first product from its Superintelligence Labs, marking a pivot away from its open-weights strategy toward a closed model. Muse Spark is a natively multimodal reasoning model supporting tool use and multi-agent orchestration, with three reasoning modes and a novel 'thought compression' post-training technique using RL to penalize excessive reasoning tokens. The model ranks fourth on the Artificial Analysis Intelligence Index and matches Llama 4 Maverick's capabilities with over an order of magnitude less training compute, though it trails in coding and agentic benchmarks. The issue also covers broader industry themes including AI-native software engineering team structures, big pharma AI adoption, and regulatory developments.