Meta launches Llama 3.1 405B: the new open-source AI model redefining innovation | Meta Facebook | Whatsapp Meta Download | Meta Whatsapp | Turtles AI
Meta launches Llama 3.1 405B: the new open-source AI model redefining innovation
With multilingual support and advanced features, Llama 3.1 opens new possibilities for developers and researchers
Highlights:
- Meta introduces Llama 3.1 405B, an advanced open-source AI model.
- Llama 3.1 supports a 128K context and eight languages.
- Collaborations with over 25 technology partners support the launch.
- Introduction of security tools like Llama Guard 3 and Prompt Guard.
Meta launches Llama 3.1 405B, a milestone in open source AI, set to redefine technological innovation with advanced features and multilingual support.
Meta has enthusiastically announced the introduction of the Llama 3.1 405B model, the most advanced among available open-source AI models. In a letter from Mark Zuckerberg, the value of open source for developers, Meta, and society is explained. With an extended context up to 128K and support for eight languages, this model promises unprecedented flexibility, control, and the capability to generate synthetic data and distill models, opening new possibilities for innovation.
The Llama 3.1 405B model, defined as a frontier model, offers unparalleled versatility and state-of-the-art capabilities, rivaling the best closed models. Meta continues to develop the Llama ecosystem, integrating additional components, including a reference system, security tools like Llama Guard 3 and Prompt Guard, and the new Llama Stack API to facilitate use by third-party projects.
Meta has partnered with over 25 partners, including AWS, NVIDIA, Databricks, Dell, Azure, and Google Cloud, ensuring services from the first day of launch. The model can be tested in the United States via WhatsApp and on the meta.ai website by posing complex math or coding questions.
The implementation of the Llama 3.1 405B model includes an iterative post-training process, with improvements in the quality of synthetic data used for Supervised Fine-Tuning (SFT), Direct Preference Optimization (DPO), and advanced filtering techniques to ensure the highest data quality. Meta has adopted a modular development approach, with components that can be used to create custom agents and new agentic behaviors.
The availability of Llama 3.1 models on platforms like Hugging Face and meta.ai allows the developer community to immediately leverage the model’s advanced capabilities, with support from technology partners to optimize large-scale inference and fine-tuning.
Llama 3.1 has been evaluated on over 150 benchmark datasets, showing competitive performance with leading models like GPT-4 and Claude 3.5 Sonnet. Updated versions of the 8B and 70B models offer significant improvements in multilingual translation, advanced tool use, and reasoning capabilities.
The Llama 3.1 405B model uses a transformer decoder-only architecture, with a post-training process that continuously improves synthetic data quality through multiple optimization cycles. The 8-bit quantization reduces computational requirements, making the model executable on a single server node.
Meta has introduced security tools like Llama Guard 3 and Prompt Guard and has initiated a request for comments on the Llama Stack API on GitHub to facilitate interoperability among ecosystem components. Security has been further strengthened through risk discovery exercises and specific enhancements for responsible AI use.
The Llama ecosystem, supported by key partners like AWS, NVIDIA, and Databricks, offers low-latency inference solutions and optimizations for cloud and on-prem implementations. Meta has collaborated with community projects like vLLM, TensorRT, and PyTorch to ensure support from day one of release, promoting further research and developments in model distillation.