Meta Llama (2023-2025): the open-source frontier-model bet that reshaped the AI competitive landscape
Meta released the first version of Llama in February 2023 with weights available to researchers under restricted academic license. In July 2023 Meta released Llama 2 in partnership with Microsoft with weights freely available for most commercial use. In April 2024 Meta released Llama 3 (8B and 70B parameter variants), followed by Llama 3.1 in July 2024 with a 405B-parameter model and substantially expanded multilingual and tool-use capabilities. The strategic bet was structurally different from OpenAI, Anthropic, and Google: Meta would make frontier models freely available rather than monetise through API or subscription. The thesis: by becoming the default open-source frontier-model provider, Meta would shape AI infrastructure in ways that constrain competitor moats while supporting Meta's own internal AI applications. By mid-2024 Llama usage had grown ~10x since 2023, with Meta claiming Llama as the leading engine of open-source AI innovation. The case is the defining current example of open-source competition against proprietary-AI-platform incumbents.
- Story: Meta released Llama AI models as open-source starting Llama 2 (July 2023). Llama 3 (April 2024), Llama 3.1 (July 2024) increasingly capable. Strategic differentiation from closed AI by OpenAI, Google, Anthropic. Mark Zuckerberg multiple public commitments to open-source AI.
- Why it matters: Meta Llama is a defining open-source AI strategy case — demonstrating that major incumbent can use open-source as competitive lever against closed-source competitors.
- Takeaway: Open-source can build developer ecosystems closed-source alternatives don't match.
- Takeaway: Major incumbent can use open-source as competitive lever against closed-source competitors.
- Takeaway: Open-source AI strategy requires sustained infrastructure investment over years.
Meta Llama AI — the four-step story
Meta Llama by the numbers
Quick facts
Where Meta AI was before 2023
Meta had been investing heavily in AI research for years through Facebook AI Research (FAIR), founded in 2013 under Yann LeCun's direction. The research output was substantial — PyTorch (the dominant ML framework outside of Google's TensorFlow ecosystem) was a Meta-built open-source project; Meta researchers had published foundational papers on transformers, self-supervised learning, and other AI topics. But Meta had not commercialised AI products at the scale Google had with Bard, Microsoft was about to with OpenAI integration, or OpenAI had with ChatGPT. The November 2022 ChatGPT launch and the rapid 2023 AI-product-race meant Meta needed a defined strategy for frontier models.
The strategic question for Meta was specifically how to compete given that Google, Microsoft, and OpenAI had multi-billion-dollar lead positions in foundation-model commercialisation. The default option (compete proprietary-AI-platform-to-proprietary-AI-platform) would have required Meta to invest enormous capital catching up on a customer-facing product layer (Meta does not have a developer-platform business comparable to AWS, Azure, or Google Cloud). The alternative option (open-source the models freely, monetise through Meta's existing consumer-products business) was the strategic bet Meta ultimately made.
The Llama release sequence (2023-2024)
Meta released Llama 1 on February 24, 2023. The initial release was research-only license with weights available via application process to academics and organizations in government, civil society, and academia. The weights leaked online within weeks of release; Meta's response was relatively muted, reflecting that the open-source-leaning strategy was already taking shape. Llama 1 (7B-65B parameter variants) was meaningfully behind GPT-4 in capability but established the open-weight-frontier-model competitive position.
On July 18, 2023 Meta released Llama 2 in partnership with Microsoft, with weights freely available for most commercial use (with restrictions for very large companies). Llama 2 (7B-70B parameter variants) was trained on 40 percent more data than Llama 1 and was the first time a frontier-capable model had been broadly licensed for commercial use. The release was a substantial strategic statement: Meta was committing to open-source competition against OpenAI and Anthropic. In April 2024 Llama 3 (8B and 70B) released with substantial capability improvements. In July 2024 Llama 3.1 added a 405B-parameter model that matched or exceeded GPT-4-class capabilities on many benchmarks — the first openly available model at that capability level.
The strategic effect and 2024-2025 trajectory
The strategic effect of the Llama release sequence has been substantial. Cloud providers (AWS Bedrock, Azure AI, IBM watsonx, Databricks, Together AI, Anyscale, others) all offer Llama-based hosted offerings, often at substantial cost advantage to proprietary frontier models. Enterprise customers concerned about vendor lock-in have moved meaningful workloads to Llama-based deployments. Researchers and academics build research on Llama because the weights are accessible in ways OpenAI and Anthropic models are not.
By mid-2024 Meta reported Llama usage had grown approximately 10x since the 2023 baseline. The model has been the engine of substantial open-source AI innovation including specialised fine-tunes (Code Llama for programming, various domain-specific variants), tool integrations, and downstream products. Meta itself uses Llama internally for the Meta AI assistant in WhatsApp, Instagram, and Facebook, plus advertising and recommendation systems. The strategic flywheel: open-source positioning produces ecosystem influence and developer mindshare; internal use of the same models produces Meta product capabilities; the combination supports Meta's broader AI ambitions without requiring a separate developer-platform monetisation business.
How RGM thinks about open-source competition against proprietary platforms
When clients ask about open-source competitive strategy against established proprietary platforms, the Meta Llama case is the defining recent example of how open-source positioning can shift competitive dynamics in a category. Three structural lessons. First, the open-source bet only works when the company has alternative monetisation that does not depend on the open-source product directly. Meta makes money from advertising in WhatsApp, Instagram, and Facebook; it can give Llama away because it does not need Llama to be a revenue line. Second, the strategic effect compounds across the broader ecosystem — every developer who builds on Llama produces evidence that proprietary alternatives are not the only path, which constrains OpenAI and Anthropic's ability to charge for what they offer. Third, the open-source positioning is specifically threatening to incumbents whose moat depends on model exclusivity — it forces them to compete on dimensions (developer experience, fine-tuning capabilities, enterprise features) that are harder to defend than pure model capability.
The pattern is hard to copy without comparable alternative monetisation. Companies whose business depends on the model layer (OpenAI, Anthropic) cannot easily open-source competitive responses because the open-source positioning would erode their own revenue. Companies with separate monetisation (Meta, Mistral via its hosting business) can sustain open-source positioning. We tell clients in foundation-model categories to be honest about whether their business model can support open-source competition or whether they need to commit to proprietary-platform strategy.
Frequently asked questions
When did Meta release Llama?
Llama 1 released February 24, 2023 (research-only license). Llama 2 released July 18, 2023 (broadly commercial license, Microsoft partnership). Llama 3 released April 18, 2024 (8B and 70B). Llama 3.1 released July 23, 2024 (8B, 70B, and 405B). Subsequent generations (Llama 3.2, 4, etc.) have continued through 2024-2025.
Is Llama really open-source?
It depends on the definition. The Llama community license allows free commercial use for most users but includes restrictions for very large companies (over 700 million monthly active users) and prohibits use for some safety-related purposes. Strict open-source advocates (Open Future, OSI definitions) argue this is not fully open-source. Pragmatic users argue the license is open-source enough for the vast majority of use cases. The model weights, code, and many training details are publicly available.
Why is Meta giving Llama away?
Strategic positioning. Meta does not have a developer-platform business comparable to AWS, Azure, or Google Cloud, so monetising frontier models through API access (the OpenAI / Anthropic / Google model) would have required Meta to build a major new business line. By open-sourcing Llama, Meta gets ecosystem influence and developer mindshare without needing to build that new business, while supporting internal Meta AI applications and constraining proprietary competitor moats.
How widely is Llama used?
Meta reported approximately 10x usage growth from 2023 to mid-2024. Cloud providers (AWS Bedrock, Azure AI, IBM watsonx, Databricks, Together AI, others) offer Llama-based hosted offerings. Many enterprise customers run Llama-based workloads. Researchers and academics build extensively on Llama because the weights are accessible. Specific usage volumes are difficult to measure precisely because much Llama use happens on self-hosted infrastructure rather than centralised cloud services.
How does Llama 3.1 405B compare to GPT-4?
On many published benchmarks, Llama 3.1 405B matches or exceeds GPT-4-class capabilities. The comparison depends on the specific benchmark and use case. Proprietary models (GPT-4 successors, Claude 3.5 Sonnet, Gemini 1.5 Pro) maintain advantages in specific dimensions including extended-context handling and multimodal capabilities. The fact that an openly-available model could match GPT-4-class capability in 2024 was the strategic milestone that made open-source frontier-model competition credible.
Sources & references
- Llama (language model) (Wikipedia) — Aggregated reference for Llama release history, model variants, and licensing.
- Introducing Meta Llama 3 (Meta AI blog) — Meta's own announcement of Llama 3 with capability and training detail.
- Introducing Llama 3.1: Our most capable models (Meta AI blog) — Meta's own announcement of Llama 3.1 including the 405B model.
- Llama 3 establishes Meta as leader in "Open" AI (IEEE Spectrum) — IEEE Spectrum analysis of Llama 3's competitive position.
- Meta releases Llama 3 open-source AI (Fortune, April 2024) — Fortune coverage of the Llama 3 release and competitive context.
- The Mirage of Open-Source AI: Analyzing Meta's Llama 2 Release Strategy (Open Future) — Critical analysis of the Llama open-source claims and licensing restrictions.