Meta AI Case Study Series · Consolidated Lineage 🦙

The Meta AI (Llama) Lineage: A Consolidated Comparison

A sixth lab, and the first whose defining choice is a distribution strategy rather than a single model: since February 2023, Meta has published Llama's weights more openly than any other frontier lab in this project, then spent 2025 openly reconsidering that bet after a benchmark-gaming scandal, a stalled flagship model, and a copyright lawsuit. This article covers the full lineage — Llama 1 through Llama 4's Scout, Maverick, and the still-unreleased Behemoth — the open-weight license that a standards body says isn't actually open source, the LMArena controversy, the Kadrey v. Meta copyright case, and the 2025 reorganization into Meta Superintelligence Labs — closing with the first six-way comparison across every lab this site has documented.

FrontierAGI Team

September 16, 2026 · 80 min read

Opening Full Timeline The Founding Llama 1 Llama 2 Llama 3 Family Llama 4 Open Weights, Not Open Source Controversies The MSL Reorg Cross-Cutting Threads Six Labs, Compared What Stayed Constant Readiness Checklist What's Missing References Videos Verdict

The Lab That Bet on Openness, Then Second-Guessed It

Every lab this project has covered eventually made a defining bet: OpenAI bet on scale, Anthropic bet on safety-first pacing, Google DeepMind bet on integration, xAI bet on speed, DeepSeek bet on efficiency. Meta's bet was distribution — release capable model weights openly enough that they become the default substrate other companies, researchers, and even rival labs build on. That bet mostly worked for three years. Then, in 2025, a chat-tuned Llama 4 variant quietly submitted to a public benchmark was caught being materially different from the model everyone else could download, the largest planned model in the family stalled and was never released, and Meta's own reporting suggests the company began weighing a partial retreat from full openness at the frontier. This article treats that arc honestly — Meta's open-weight releases are real and were genuinely influential, and the retreat from full openness is also real and not yet resolved.

The Full Timeline

From FAIR to Four Llama Generations to a Reorg

2013FAIR founded

Feb 2023Llama 1

Jul 2023Llama 2

Dec 2023Purple Llama

Apr 2024Llama 3

Jul 2024Llama 3.1 (405B)

Sep 2024Llama 3.2

Dec 2024Llama 3.3

Apr 2025Llama 4 + LMArena controversy

May–Jun 2025Behemoth delayed

Jun 2025Kadrey v. Meta summary judgment

Jul 2025Meta refuses EU AI Act code

Aug 2025Meta Superintelligence Labs formed

Nov 2025Yann LeCun departs

Amber = the two events that most directly damaged Meta's open-weights story: the LMArena scandal and Behemoth's stall.

$14.3B Meta's investment in Scale AI to bring in Alexandr Wang as Chief AI Officer, one of the most expensive talent acquisitions in tech history

32nd The public, released Llama 4 Maverick's actual LMArena rank, once the undisclosed chat-tuned variant was excluded

700M Monthly active user threshold above which Meta's own license requires a company to separately request permission to use Llama

The Founding

From FAIR to GenAI to Meta Superintelligence Labs

Meta's AI research traces back to Facebook AI Research (FAIR), founded in 2013 under Yann LeCun, who spent roughly twelve years as the company's Chief AI Scientist before departing in November 2025. Unlike DeepMind or OpenAI, Meta's Llama models were not FAIR's direct product — they came out of a separate "GenAI" organization built specifically to ship consumer- and developer-facing generative AI products, while FAIR continued longer-horizon fundamental research. That split, and the tension between shipping speed and research depth it implied, ran through the entire Llama era and became the explicit subject of Meta's 2025 reorganization (covered in a dedicated section below).

The Founding Idea Split fundamental AI research (FAIR) from a dedicated generative-AI product organization (GenAI) tasked with shipping openly-released, broadly adopted model weights — treating distribution, not just capability, as the primary strategic lever.

Llama 1

Compute-Efficient Small Models, Leaked Within a Week

Llama 1Feb 2023

Research Trigger: Test the Chinchilla compute-optimal scaling thesis directly — train smaller dense models on far more tokens than prior practice, aiming to beat larger, under-trained models at inference time rather than chase parameter count.
Network Architecture: Dense transformer at four sizes — 7B, 13B, 33B, and 65B parameters.
Data: Trained entirely on publicly available data (no proprietary or licensed corpora) — roughly 1 trillion tokens for the 7B/13B models and roughly 1.4 trillion tokens for the 33B/65B models, drawn from sources including CommonCrawl, C4, GitHub, Wikipedia, books, and ArXiv.
Influence on Next: Despite a research-only, gated, non-commercial license, the weights leaked publicly within about a week of announcement, spreading uncontrolled across the open community — the direct trigger for Llama 2's shift to a genuinely commercial license seven months later, an attempt to get ahead of unauthorized redistribution rather than fight it.

Meta's own reported benchmarks claimed Llama 1's 13B model outperformed GPT-3's 175B model on most evaluated benchmarks, and the 65B model was competitive with Chinchilla-70B and PaLM-540B — a genuinely disclosed, testable claim about training efficiency, three years before DeepSeek made a similar claim the center of a much larger controversy.

Llama 2

The First Truly Commercial Open-Weight Frontier Model

Llama 2 & Llama 2-ChatJul 2023

Network Architecture: Dense transformer at 7B, 13B, and 70B (a 34B variant was trained but held back from release pending further safety review); context length doubled to 4,096 tokens.
Training Technique: RLHF pipeline built around a reward model with a scalar regression head trained on human preference comparisons, combining a modified binary ranking loss with rejection-sampling fine-tuning (used for the 70B model) and PPO balancing separate helpfulness and safety reward signals.
Learning Technique & Scaffolding: Ghost Attention (GAtt), a context-distillation-derived technique that keeps a system-prompt-level instruction present across many conversational turns, fixing a common multi-turn instruction-forgetting failure.
Safety Framework: Iterative red-teaming across model versions, evaluated with human ratings on roughly 2,000 adversarial prompts spanning risk categories (dangerous activities, hate speech, unqualified professional advice) in both single-turn and multi-turn settings.

Llama 2's license was the first in the family to permit genuine commercial use, distributed alongside a partnership placing the models on Microsoft Azure — a distribution choice that made Llama 2 the first realistic "download a frontier-adjacent model and ship a product" option most companies had.

Llama 3 Family

Four Releases in Nine Months

Apr 2024Llama 3

8B and 70B: a new, larger tokenizer vocabulary and Grouped-Query Attention (GQA) across all sizes for inference efficiency, trained on a substantially larger corpus than Llama 2.

Jul 2024Llama 3.1

Adds a 405B dense model — the largest openly released dense model at the time. Trained on roughly 15.6 trillion tokens using more than 16,000 Nvidia H100 GPUs, with pretraining compute on the order of 3.8×10²⁵ FLOPs; context extended to 128K tokens via continued training after an initial 8K-context pretraining phase; natively multilingual across eight languages. Published as "The Llama 3 Herd of Models" (arXiv:2407.21783). This release also introduced the 700-million-monthly-active-user licensing threshold discussed in the licensing section below.

Sep 2024Llama 3.2

First multimodal Llamas plus on-device sizes: vision-capable variants alongside small, text-only 1B and 3B models explicitly designed to run on-device rather than in a datacenter — Meta's first explicit move toward edge deployment rather than server-only inference.

Dec 2024Llama 3.3

A 70B model reported to match 405B-level performance on many benchmarks — an efficiency-focused release closing out the Llama 3 family by making frontier-adjacent quality available at a much smaller, cheaper-to-serve size.

Llama 4

The First MoE Llama, and a Benchmark Scandal

Llama 4: Scout, Maverick & BehemothApr 2025

Network Architecture: Meta's first Mixture-of-Experts Llama family, with native multimodality built in from pretraining rather than added afterward. Scout was marketed with an unusually large 10-million-token context window claim.
Influence on Next: Behemoth, the largest model in the family, was originally planned to launch alongside Scout and Maverick at Meta's first LlamaCon developer conference, then pushed to June 2025, then to fall 2025 or later. As of the most recent available reporting it remains unreleased and undated — not formally cancelled, but effectively shelved after reported mid-training issues with MoE routing and chunked attention at roughly 2-trillion-parameter scale; Meta reportedly used it internally as a "teacher model" to help train Scout and Maverick via distillation instead.

The LMArena Controversy Meta submitted a variant called "Llama-4-Maverick-03-26-Experimental" — described as optimized for conversational style — to the LMArena leaderboard, where it scored an ELO of 1417, ranking #2 and above GPT-4o. This was not the model the public could actually download. Once the discrepancy surfaced, the publicly released "vanilla" Maverick's real LMArena rank turned out to be 32nd. LMArena responded that "Meta's interpretation of our policy did not match what we expect from model providers" and changed its submission policy to prevent undisclosed custom-tuned variants going forward. In January 2026, Yann LeCun — by then departed from Meta — told the Financial Times the results "were fudged a little bit."

Open Weights, Not Open Source

The Most Prominent Open-Weight Strategy at Frontier Scale — With a Catch

Of every lab this project has covered, Meta has released the most capable models with the most permissive-sounding licenses most consistently — GPT-1 and Grok-1 were each open-weighted once; DeepSeek released V3 and R1 openly; Meta has done this across four full generations. But the Llama Community License is explicitly not an OSI-approved open-source license. It layers an Acceptable Use Policy, attribution and naming requirements, and — starting with Llama 3.1 — a hard commercial restriction on top of the free grant.

The 700M-MAU Clause Any company or product with more than 700 million monthly active users at the time of a given Llama release cannot use the default free license — it must separately request a license from Meta. The clause is widely understood to be aimed squarely at Meta's largest competitors (Google, Microsoft, Amazon, and similarly scaled platforms), not at ordinary developers or startups.

The Open Source Initiative's own review, and a separate Free Software Foundation evaluation, both concluded the license fails core open-source tests: it restricts Freedom 0 (use for any purpose), discriminates against certain classes of user (later license versions reportedly excluded persons located in the EU from some capabilities), and restricts fields of endeavor through the Acceptable Use Policy and the MAU cap. Meta's own marketing language calls Llama "open source"; the standards bodies that define that term disagree. This project uses "open weight" rather than "open source" for Llama throughout, for the same reason it does for DeepSeek and Grok-1.

Controversies

A Copyright Case, and a Refused EU Code

Filed 2023 · Ruling Jun 2025Kadrey v. Meta

Thirteen published authors — including Sarah Silverman, Ta-Nehisi Coates, and Junot Díaz — sued Meta over training Llama on their copyrighted books without permission. Unsealed court filings reported that Meta accessed the Library Genesis (LibGen) database, a large repository of pirated books, starting in October 2022 to evaluate its training value, and after failing to secure licenses, proceeded to train on LibGen-sourced text in spring 2023 following escalation to Mark Zuckerberg himself. In June 2025, Judge Vince Chhabria granted Meta summary judgment on the plaintiffs' "market harm" theory, ruling that specific use qualified as fair use — but reporting into 2026 indicates the court allowed plaintiffs to amend their complaint with a contributory-infringement theory, and the case remains ongoing.

Jul 2025Refusing the EU AI Act Code of Practice

Meta's Chief Global Affairs Officer, Joel Kaplan, announced Meta would not sign the EU's voluntary Code of Practice for the AI Act, calling it "over-reach" that "introduces legal uncertainties" beyond the Act's own scope. The code — drafted by independent experts and requiring signatories to publish training-data summaries and affirm EU copyright compliance — was signed by OpenAI and Mistral; Meta's refusal made it the most prominent named holdout among major labs.

The MSL Reorg

Meta Superintelligence Labs, and the Departure of Yann LeCun

In mid-2025, Meta invested $14.3 billion to bring in Scale AI founder Alexandr Wang as Chief AI Officer — one of the most expensive individual talent acquisitions in the industry's history. On August 19, 2025, Meta formalized a reorganization into Meta Superintelligence Labs (MSL), split into four units: TBD Lab (led directly by Wang, focused on next-generation LLMs powering the Meta AI assistant), FAIR (now led by Rob Fergus, focused on long-term fundamental research), Products and Applied Research, and MSL Infra (compute and datacenter operations). Yann LeCun, after roughly twelve years as Chief AI Scientist, departed Meta in November 2025 following the restructuring — he was reportedly asked to report to Wang, and publicly described Wang afterward as "young and inexperienced," warning that other researchers were leaving as well.

Reported, Not Yet Confirmed by Meta Directly Multiple reports through late 2025 and into 2026 describe Meta shifting toward a more proprietary model strategy — including a reportedly closed-weight project under MSL and a next open-weight "Llama 5" release — as a direct response to Llama 4's stumbles and competitive pressure from open releases like DeepSeek's. These reports come from secondary and aggregator sources rather than a Meta primary announcement, and this article treats the exact model names and specifications involved as unconfirmed while treating the broader strategic reconsideration as well-corroborated across independent reporting.

Cross-Cutting Threads

Three Patterns Unique to This Lineage

Thread 1 — Distribution as the Core Bet, Not Just Capability Every other lab in this project competes primarily on capability or safety posture. Meta competed primarily on making capable weights available to the widest possible set of downstream builders — a strategic axis distinct from any other lab's defining choice.

Thread 2 — The Only Lab Caught Gaming a Public Benchmark The LMArena incident is the only documented case in this project of a lab submitting a materially different, undisclosed model variant to a public leaderboard than the one it actually shipped — a distinct category of integrity failure from Grok's live safety incident or DeepSeek's disputed cost claim.

Thread 3 — Openness in Public Retreat, in Real Time Meta is the only lab in this project whose commitment to its own defining strategy — open weights — visibly wavered within the period this article covers, with credible reporting of a pivot toward more closed development following Llama 4's reception.

Six Labs, Compared

OpenAI, Anthropic, Google DeepMind, xAI, DeepSeek, and Meta

Dimension	OpenAI	Anthropic	Google DeepMind	xAI	DeepSeek	Meta
Origin story	A research paper (2018)	A safety-pace disagreement (2021)	A corporate merger (2023)	A founder-led startup (2023)	A hedge fund's AI research spinoff (2023)	A social media company's research lab (2013), later split into FAIR + GenAI + MSL
Architecture disclosure	Open, then closed	Undisclosed from day one	Partial (e.g., confirmed MoE)	Open once (Grok-1), closed since	Detailed and disclosed (MLA, MoE routing, exact params)	Open weights across four generations, but under a non-OSI license
Compute disclosure	Precise (GPT-1, GPT-3), then none	Never disclosed	Never disclosed	Never disclosed	Disclosed, but disputed and contested by outside analysis	Precisely disclosed for Llama 3.1 (GPU count, FLOPs), silent since
Safety framework type	Voluntary capability-risk framework	Voluntary capability-risk framework	Voluntary capability-risk framework	Voluntary framework, publicly criticized as weak	State-mandated political-content compliance regime	Voluntary Frontier AI Framework (Feb 2025), paired with Purple Llama's Llama Guard content classifiers
Most consequential public event	GPT-6 Astra's "Critical" classification	Opus 4/Sonnet 4's ASL-3 activation	No CCL reached to date	The MechaHitler incident	A $600B single-day market shock	The LMArena benchmark-gaming scandal
National government bans	None documented	None documented	None documented	None documented	Italy (app store removal), Australia (government devices)	None documented; refused to sign the EU AI Act's voluntary Code of Practice

What Stayed Constant

The Same Foundation, a Sixth Time

Llama 4's Mixture-of-Experts architecture is, once again, a variation on the same 2017 Transformer design every model in this six-lab project shares. Six labs, six founding stories, six disclosure philosophies, six different relationships with openness and government oversight — and underneath every one of them, including the lab that built its entire public identity around giving its weights away, the same architectural family.

Readiness Checklist

Can you explain why the Llama Community License is described as "open weight" rather than "open source" in this article, and name the specific clause a standards body objected to?

Can you explain what actually happened in the LMArena controversy, and why it's a different kind of integrity failure than DeepSeek's cost dispute or Grok's MechaHitler incident?

Can you name what changed structurally at Meta in August 2025, and who left as a result?

Could you explain, to someone who only remembers "Meta open-sources its AI," why that framing is not quite accurate?

⚠️ What's Missing or Uncertain

Everything dated after this article's own January 2026 research cutoff should be treated with extra caution. Reports of a next open-weight "Llama 5" release, a separate closed-weight Meta model, and an updated version of Meta's Frontier AI Framework circulated in late 2025 and into 2026, but come primarily from secondary and aggregator sources rather than Meta's own official channels or major wire services — this article does not treat their specific names, sizes, or feature claims as settled fact. Behemoth's exact status (indefinitely delayed vs. quietly cancelled) is also not confirmed directly by Meta. The Kadrey v. Meta litigation remains ongoing as of the most recent available reporting and its ultimate outcome is not yet known.

🔗 Reference Links

🎥 Recommended Videos

But What Is a Neural Network?

3Blue1Brown — foundational context for the architecture family every Llama model, MoE included, still belongs to

Attention Is All You Need — Paper Explained

Yannic Kilcher — the 2017 Transformer paper underlying every model across all six labs this project has documented

🧭 Closing — Openness Was Real, and So Was the Retreat From It

🎯 The Bottom Line

Meta's open-weight releases genuinely reshaped the field — Llama 2 and Llama 3.1's 405B model gave more developers, researchers, and even competing labs a capable base model than any other lab's releases in this project. But 2025 exposed real costs to that strategy that this article does not soften: a benchmark-gaming scandal that damaged trust in Meta's own reported results, a flagship model (Behemoth) that never shipped, a copyright lawsuit over how the training data was sourced, and credible reporting that the company itself is now reconsidering how open its next generation of frontier models will be. Distribution was Meta's defining bet — and unlike every other lab's defining bet in this project, this is the first one this project has documented the lab itself visibly walking back.