Llama: the model that leaked, started the open-weights movement, and then changed course

LLaMA - Large Language Model Meta AI - was released on February 24, 2023, and only to researchers who applied for access. It stayed private for about a week. On March 3 the weights appeared in a torrent shared on 4chan, and within days people were running a GPT-3-class model on their own hardware. Meta did not plan to start the open-weights movement, but that leak arguably did.

What followed was a burst of creativity. Stanford’s Alpaca fine-tuned the 7B model for under $600. Georgi Gerganov wrote llama.cpp, a plain C/C++ inference engine that ran LLaMA on a MacBook CPU - today it has more than 127,000 GitHub stars and gave the world the GGUF format. Ollama, which started as a friendly wrapper around it, has even more: over 180,000 stars. Local AI as we know it grew out of a leaked research model.

Meta leaned in. Llama 2, in July 2023, came with a license that allowed commercial use - with a famous catch: companies with more than 700 million monthly active users needed Meta’s permission, which in practice meant a handful of competitors. Llama 3 was trained on about 15 trillion tokens, and Llama 3.1 in July 2024 shipped a 405B model with a 128K context window - the first open-weights model that could stand next to the frontier closed ones. Llama 3.2 added 1B and 3B models small enough for phones. By early 2025 Meta was reporting more than a billion downloads.

“Open source” was always an argument, though. The Open Source Initiative never accepted the Llama license as open source because of its usage restrictions, and the Free Software Foundation classified the Llama 3.1 license as nonfree. The accurate term is open weights: you can download, run, and fine-tune the model, but the license, not you, sets the rules. For a business that difference matters - read the acceptable use policy before you build on it.

Llama 4 was released on a Saturday - April 5, 2025 - and switched to a mixture-of-experts architecture. Scout has 17B active parameters out of 109B and a 10-million-token context window; Maverick has 17B active out of 400B. The launch was overshadowed by a benchmark controversy: the version that ranked high on LMArena was an experimental chat-optimized variant, not the one Meta released. The giant Behemoth, about 2 trillion parameters, was announced and never shipped. In April 2026 Meta Superintelligence Labs launched Muse Spark - a closed, proprietary model. Meta says existing Llama models stay available and has talked about open-sourcing future versions, but the era of an open-weights Meta flagship looks paused.

For builders the lesson is practical. Llama made it normal to run a capable model on your own servers - for privacy, cost control, or offline use - and that option is not going away, whoever leads open weights next. In a TypeScript or Python product I treat the model as a replaceable dependency: an OpenAI-compatible endpoint whether it is a hosted API, vLLM, or Ollama on a local box, plus evaluations that tell me when a cheaper open model is good enough. The winning architecture is the one where switching models is a config change.