Llama

On ToTheMoon, Llama is the channel's standing example of how an open model becomes a large company's strategy. The hosts come back to it whenever they weigh what openness actually buys: a way to catch up with the leader, a distribution channel or a political lever. The central point, set out in episode 053, is that Llama's value to Meta lies not in model size but in the right to embed it anywhere.

In 2024 Llama was treated as a working tool. In episode 003 the hosts mentioned that Llama 3 matched GPT-4 on a number of benchmarks and spoke from their own video-generation projects, where Llama 2 had not been good enough but Llama 3 was. Episode 017 looked at how Meta was folding Llama into WhatsApp and Facebook to outrun ChatGPT on audience. The other side of openness surfaced in episode 032, when the hosts noted that the Chinese military had built its own solution on top of Llama.

In spring 2025 Llama got a dedicated treatment. Episode 052 read a billion downloads as a sign that reasoning was no longer the exclusive preserve of one company, while noting that DeepSeek's models were more open than Llama and that Meta had fallen behind. In episode 053 the two trillion parameters of Llama 4 Behemoth were dismissed as a number that says little about outcomes; what matters is whether the system can be fine-tuned and run next to one's own data. Episode extra03 unpacked the community license, which looks open but limits very large services. Episode 088 saw Meta trying to return with a new Llama, and in episode 141 Llama produced odd results in a study of political restrictions.