MANILA, July 12 — When Meta released LLaMA 2 to the public in July 2023, it did something unusual in the world of big tech: it gave the model away for free. Two years on, that decision has reshaped how developers and researchers across Asia are building with artificial intelligence. The open-source large language model, trained on 2 trillion tokens of public data, has been downloaded more than 100 million times as of January 2025.
For startups in the Philippines, India, and Vietnam that cannot afford the steep per-token fees of proprietary models, this has been a game changer. LLaMA 2 comes in three sizes — 7 billion, 13 billion, and 70 billion parameters — meaning a small team with modest computing resources can still work with a capable AI.
To understand the scale of the cost difference, one only needs to look at the competition. OpenAI’s GPT-4, released in March 2023, charges $0.03 per 1,000 tokens for API access. Google’s Gemini, which arrived in December 2023, is also proprietary.
For a developer in Manila building a customer-service chatbot for a local cooperative, those costs add up fast. Meta’s approach removes that barrier entirely.
An ecosystem of derivative models
The impact has been measurable. On platforms like Hugging Face, more than 10,000 derivative models have been created from LLaMA 2. Researchers have fine-tuned it for Tagalog, Bahasa Indonesia, and Vietnamese.
Startups have adapted it for agricultural advice, local-language education tools, and medical triage. Closer to home, this means AI tools that actually understand the languages and contexts of communities across Southeast Asia — not just English.
Meta did not stop with LLaMA 2. In February 2025, the company announced LLaMA 3, claiming it outperforms GPT-4 on several benchmarks, including the Massive Multitask Language Understanding test and HumanEval for code generation. If those claims hold up in independent testing, the implications for developers in the region are significant: a state-of-the-art model, free to use, that can write code and answer complex questions across multiple domains.
What this means for communities
This strategy positions Meta as a leader in open-source AI. But it also raises questions about influence. By making LLaMA freely available, Meta gains a foothold in the tools that thousands of startups and researchers build their products on.
For families and small businesses across Asia that rely on these technologies, the choice between open and proprietary models is not just about cost — it is about who controls the infrastructure underneath. For now, the open path has won many converts.
The 100 million downloads and 10,000 derivative models are proof that developers vote with their code. As LLaMA 3 rolls out and competes with GPT-4 on performance, the conversation for many in the region will shift from “Can we afford AI?” to “What can we build with it?”
