Meta’s research team introduced Llama 3.1 on July 23, 2023, calling it “the world’s largest and most capable openly available foundation model.” Llama 3.1 is available in various parameter sizes — 8B, 70B, and 405B — providing flexibility for deployment based on computational resources and specific application needs. On April 18, 2024, Meta announced the Llama 3 family of large language models, which initially included only the 8B and 70B sizes. This latest release introduced the 405B model along with upgraded versions of the 8B and 70B models. Llama 3.1 models represent a significant advancement over their predecessor, Llama 2, being pre-trained on an extensive corpus of 15 trillion multilingual tokens, a substantial increase from Llama 2’s 1.8 trillion tokens. With a context window of up to 128k tokens — previously limited to 8k tokes — they offer notable improvements in multilinguality, coding, reasoning, and tool usage. Llama 3.1 maintains a similar architecture to Llama and Llama 2 but achieves performance improvements through enhanced data quality, diversity, and increased training scale. Meta’s research team tested Llama 3.1 on over 150 benchmark datasets covering a wide range of languages. They found that their “flagship model” with 405B parameters is competitive with leading models across various tasks and is close to matching the state-of-the-art performance. The smaller models are also “best-in-class,” outperforming alternative models with comparable numbers of parameters. In multilingual tasks, the small Llama 3.1 8B model surpassed Gemma 2 9B and Mistral 7B, while Llama 3.1 70B outperformed Mixtral 8Xx22B and GPT 3.5 Turbo. Llama 3.1 405B is on par with Claude 3.5 Sonnet and outperformed GPT-4 and GPT 4o. Meta’s research team emphasized that Llama 3.1 405B is “the first openly available model that rivals the top AI models when it comes to state-of-the-art capabilities in […] multilingual translation,” among other tasks. They expressed optimism about the potential for creating innovative applications leveraging the model’s multilingual capabilities and extended context length, stating, “we can’t wait to see what the community does with this work.”. In addition to language processing, the development of Llama 3.1 included multimodal extensions that enable image recognition, video recognition, and speech understanding capabilities. Although these multimodal extensions are still under development, initial results indicate competitive performance in image, video, and speech tasks. Meta’s research team specifically evaluated Llama 3.1 on automatic speech recognition (ASR) and speech translation. In ASR, they compared its performance against Whisper, SeamlessM4T, and Gemini. Llama 3.1 outperformed Whisper and SeamlessM4T across all benchmarks and performed similarly to Gemini, demonstrating “strong performance on speech recognition tasks.” In speech translation tasks, where the model was asked to translate non-English speech into English text, Llama 3.1 again outperformed Whisper and SeamlesM4T. “The performance of our models in speech translation highlights the advantages of multimodal foundation models for tasks such as speech translation,” Meta’s team said. They also shared details of the development process to help the research community understand the key factors of multimodal foundation model development and encourage informed discussions about the future of these models. “We hope sharing our results early will accelerate research in this direction,” they said. Meta’s launch of Llama 3.1 has created a buzz in the AI community. Since the release, many people have taken to X and LinkedIn to call it a “game-changer” or “GPT-4 killer,” recognizing this moment as “the biggest moment for open-source AI.” Additionally, they have talked about a “seismic shift in business transformation,” explaining that this is going to “revolutionize how companies work.” Posts are filled with examples showing the many different ways Llama 3.1 can be used, building from phone assistants to document assistants and code assistants. Meta has released all Llama 3.1 models under an updated community license, promoting further innovation and responsible development towards artificial general intelligence (AGI). “We hope that the open release of a flagship model will spur a wave of innovation in the research community, and accelerate a responsible path towards the development of artificial general intelligence” they said. Additionally, they believe that the release of Llama 3.1 will encourage the industry to adopt open and responsible practices in AGI development. The Meta research team acknowledges that there is still much to explore, including more device-friendly sizes, additional modalities, and further investment in the agent platform layer. The models are available for download on llama.meta.com and Hugging Face and ready for immediate development within a broad ecosystem of partner platforms, including AWS, NVIDIA, Databricks, Groq, Dell, Azure, Google Cloud, and Snowflake. Ahmad Al-Dahle, who leads Meta’s generative AI efforts, wrote in a post on X, “With Llama 3.1 in NVIDIA AI Foundry we’ll see enterprises to easily create custom AI services with the world’s best open source AI models.”