# Open Source AI is the Path Forward

Mark Zuckerberg published this letter on Meta's newsroom on 23 July 2024, alongside Llama 3.1 405B, which he called the first frontier-level open source AI model, and updated 70B and 8B versions. It is the clearest statement of why Meta released Llama, which is why Nathan Lambert put it on his [[open-source-ai-reading-list]]. The argument has three parts, for developers, for Meta, and for the world, and opens with an analogy: the big companies of the high-performance computing era each built their own closed Unix, and Linux beat them all, first on price and freedom to modify, later on capability and security.

Like most writing on the topic, the letter says "open source" and means open weights. Licensing and training data are not discussed.

## The prediction

Zuckerberg sketches a trajectory. Llama 2 matched an older generation of models, Llama 3 was competitive with the best, and "starting next year, we expect future Llama models to become the most advanced in the industry."

That did not happen. By May 2026, Bill Gurley's [[from-open-source-software-to-open-source-strategy]] reports that Llama 4 underperformed at its 2025 launch, the frontier-class Llama 4 Behemoth was shelved, Zuckerberg said in July 2025 that superintelligence-capable models would not be released openly, and Meta Superintelligence Labs shipped a closed model, Muse Spark, in April 2026. Gurley's verdict is that the letter "reads very differently today". Lambert's [[open-models-in-perpetual-catch-up]] notes that the open models people argue about now come from China, where a year earlier they would have been Llamas. [[meta-platforms]] covers the organizational turmoil around the same turn.

## For developers

This section reports what Zuckerberg hears from developers, CEOs and officials. Organizations want to train, fine-tune and distill their own models at the size each task needs, without anyone else seeing their data. They do not want a closed provider to change the model, change the terms or stop serving them, or to be tied to the one cloud with exclusive rights. Some data cannot go to a cloud API at all. And cost: he claims Llama 3.1 405B runs on one's own infrastructure at roughly half the cost of GPT-4o. Launch partners are named to show an ecosystem already exists: Amazon, Databricks and Nvidia for fine-tuning and distillation, Groq for fast inference, every major cloud, and Scale AI, Dell and Deloitte for enterprise adoption.

## For Meta

This is the part later writers cite most. Zuckerberg's formative experience is building on Apple's platforms, with the fees, arbitrary rules and blocked products that came with it, and he does not want Meta locked into a competitor's closed AI stack. He gives four reasons that releasing Llama costs little. An ecosystem of tools, silicon optimizations and integrations only grows if others use the model; used by Meta alone, Llama would fare no better than the closed Unixes. Competition is intense enough that any single release gives away little lead. Meta does not sell model access, so openness does not cut into revenue as it would for closed providers, which he says is one reason those providers lobby governments against open source. And Meta has done this before: releasing server and data center designs through the Open Compute Project saved it billions, and PyTorch and React brought back outside improvements.

Gurley reads the Open Compute example the same way, as commoditizing an input rather than giving something away. Lambert, in [[what-comes-next-with-open-models]], doubts that "commoditize your complements" still works as a reason once a competitive model costs billions of dollars to build.

## For the world, and safety

Zuckerberg argues that open source spreads AI's benefits widely, keeps power from concentrating in a few companies, and is safer. He splits harm into unintentional and intentional. Unintentional harm, which he says covers most concerns, from bad health advice to runaway-AI scenarios, should be lower with open models because they can be scrutinized, and Llama Guard is offered as a safety layer. For intentional harm the baseline is whether a model helps more than a web search would. Against individual bad actors his answer is a balance of power: if everyone has access to similar generations of models, institutions with more compute can check those with less, the way Meta's systems catch less sophisticated attackers on its social networks.

Against states such as China, he argues that closing models would not work. Adversaries are good at espionage, weights fit on a thumb drive, and most tech companies are not secured against that. A world of closed models would leave a few large companies and geopolitical adversaries with the best models while startups, universities and small businesses go without. His alternative is a strong open ecosystem, with leading US companies working closely with the government and allies to keep a lasting first-mover advantage.

## What the letter does not engage

The safety section skips the property that separates open weights from open software: a release cannot be recalled, and a safeguard like Llama Guard is optional for anyone running the weights. The balance-of-power argument assumes defenders and attackers gain about equally from the same model generation, which is exactly what later work such as [[societal-impact-of-open-foundation-models]] and [[a-safe-path-to-open-weights]] tries to pin down. The claim that open source is "advancing at a faster rate than closed models" is presented as something "lots of people see" rather than shown, and two years of data since then put open models at a steady lag instead.

The business section has aged differently from the forecast. Zuckerberg's reasons were specific to Meta's position, and once that position changed, so did Meta's release policy. Irene Solaiman's 2023 [[gradient-of-generative-ai-release]] had already asked whether Meta's early openness with OPT would push other large companies toward open release; this letter looked like the answer, and Meta's 2026 retreat reversed it.