Llama 3
Meta's 2024 open-weight release that closed most of the gap to closed models
Llama 3 is the April 2024 release that closed much of the performance gap between Meta’s open models and closed competitors like GPT-4 and Claude 3. It shipped in two sizes, 8B and 70B, both trained on about 15 trillion tokens with an 8,192-token context window, smaller than what came just months later in Llama 3.1. At launch Meta said the 70B model beat Gemini Pro 1.5 and Claude 3 Sonnet on several benchmarks, and the 8B model outperformed the much larger Llama 2 70B on MMLU, a notable jump for a model roughly a ninth of the size. Llama 3 is largely superseded now by the 3.1, 3.2, and 3.3 generations, but it was the release that established Meta’s Llama line as a serious rival to closed frontier models.