In February 2023, Meta shipped Llama 1’s weights to a short list of approved academic researchers, not to the public. Getting access meant filling out a form and waiting for an approval email. Within a week, someone posted a working torrent link, the full model spread across the internet, and Meta never controlled that release again. There was no keynote about founding an open ecosystem, no strategy memo behind it. The company that ended up anchoring the entire open-weight movement got there because of a leak it didn’t plan, not a decision it made.

The accident that built a field

Before that leak, capable language models lived almost entirely behind company APIs. Researchers outside a handful of labs couldn’t inspect weights, fine-tune them, or run them on their own hardware. Llama 1’s unplanned spread changed the baseline overnight: within weeks, people were running a GPT-3-class model on laptops, quantizing it, fine-tuning it into dozens of variants. Meta’s response wasn’t to fight this. Five months later, in July 2023, it released Llama 2 with weights explicitly licensed for commercial use, the first time a model of that caliber came with permission built in rather than smuggled out. That single decision turned a one-time accident into a repeatable strategy, and for the next year and a half, Meta acted like a company that had chosen this path on purpose.

Two years of pushing the frontier outward

What followed reads like a deliberate campaign to keep the open frontier moving. Llama 3 arrived in April 2024. Three months later, Llama 3.1 added a 405-billion-parameter version, the first open-weight model that could genuinely be called frontier-class rather than a smaller cousin of one. September 2024 brought Llama 3.2, which pushed in the opposite direction too, adding multimodal variants alongside small models built for phones and edge devices. By December 2024, Llama 3.3 packed performance close to that 405B giant into a 70B model, a reminder that most of the previous year’s gains were less about raw parameter count and more about how the training itself had improved. Llama 4, in April 2025, moved the architecture to a mixture of experts, added native multimodality, and shipped a “Scout” variant with a 10 million token context window, larger than almost anything else on the market, open or closed.

The room after everyone else arrived

That is where the momentum stops. Since Llama 4, Meta’s newest and most capable models have increasingly stayed behind closed doors, with open releases trailing further behind the internal frontier than at any point since 2023. The pattern resembles someone who trips in public, accidentally starts a dance everyone else picks up, and two years later is standing at the edge of the floor while the party they never meant to throw carries on without them.

None of this erases what happened. The open-weight ecosystem, the fine-tuned variants, the research built on inspectable weights, all of it traces back to a release that was never supposed to leave a small group of approved researchers. The irony is specific: the wave didn’t start with a company deciding open weights were the future, and the company that triggered it now looks like the one least committed to that future. Llama’s commercial license itself was never fully open in the way “open source” implies, a distinction covered in Open Source vs Open Weight, which digs into the 700-million monthly-active-user clause buried in that license. Where the open frontier goes from here, and who picked up where Meta stepped back, is the subject of the next piece in this series.