Llama 3.1 has been leaked! You heard that right, this open-source model with 405 billion parameters has caused quite a stir on Reddit. It might be the closest open-source model to GPT-4o to date, and even surpasses it in certain aspects.
Llama 3.1 is a large language model developed by Meta (formerly Facebook). Although the official release has not yet been made, the leaked version has already stirred up the community. This model not only includes the base model but also benchmark results for 8B, 70B, and the largest 405B parameters.
Performance Comparison: Llama 3.1 vs GPT-4o
According to the leaked comparison results, even the 70B version of Llama 3.1 outperforms GPT-4o in multiple benchmarks. This marks the first time an open-source model has reached the State of the Art (SOTA) level in several benchmarks, truly showcasing the power of open-source!
Model Highlights: Multilingual Support, Richer Training Data
Llama 3.1 was trained using over 15 trillion tokens from public sources, with the pre-training data cut off in December 2023. It supports not only English but also French, German, Hindi, Italian, Portuguese, Spanish, and Thai, making it excel in multilingual conversational use cases.
The Llama 3.1 research team places a high emphasis on model security. They employed a multifaceted data collection method, combining human-generated data with synthetic data to mitigate potential security risks. Additionally, the model introduced boundary prompts and adversarial prompts to enhance data quality control.
Model card source: https://pastebin.com/9jGkYbXY#google_vignette