It seems like just yesterday that I wrote about the impact of Llama 3.1, and now we're facing another significant milestone in artificial intelligence with the release of Llama 3.2 on September 25, 2024.
Introduction
Meta has just announced Llama 3.2 , an innovation that promises to transform the landscape of edge and vision AI. This new version brings large-scale vision language models (LLMs) of small and medium size (11B and 90B) and lightweight text-only models (1B and 3B), designed to adapt to mobile and edge devices. Available in pre-trained and instruction-tuned versions, these models offer exceptional flexibility and performance for a wide range of applications.
What's New in Llama 3.2
Mobile and Edge Optimized Templates
The Llama 3.2 1B and 3B models now support a context length of up to 128K tokens , setting a new standard for applications on devices such as:
- Multilingual knowledge retrieval and summarization
- Instruction follow-up
- Rewrite tasks run locally
Optimized for Qualcomm, MediaTek, and ARM processor hardware, these models enable efficient processing without the need for heavy infrastructure.
Advances in Computer Vision
The Llama 3.2 11B and 90B vision models can immediately replace their text equivalents, outperforming even closed models like Claude 3 Haiku in image comprehension tasks. Unlike other open multimodal models, both pre-trained and aligned models are available for custom adjustments using torchtune , and can also be tested with the Meta AI assistant.
Llama Stack: Simplifying Development
Meta is launching the first official distributions of the Llama Stack , which significantly simplify how developers work with Llama templates in different environments:
- Single-node
- On-premises
- In the cloud
- On the device
This enables turnkey implementation of recovery-augmented generation (RAG) applications and security-integrated tools, accelerating development time and reducing complexity.
Strategic Partnerships
In collaboration with partners such as AWS , Databricks, Dell, Fireworks, Infosys, and Together AI , Meta is expanding the reach of the Llama Stack to enterprise customers. Distribution to devices is handled via PyTorch ExecuTorch , while single-node distribution is facilitated by Ollama.
Openness that Drives Innovation
Meta continues to share its work because it believes that openness drives innovation . Llama 3.2 leads the way in openness, modifiability, and cost efficiency, enabling more people to achieve creative and transformative breakthroughs using generative AI.
Availability
The Llama 3.2 templates are available for download at llama.com and Hugging Face , and are also accessible for immediate development across a broad ecosystem of partner platforms, including:
- AMD
- Google Cloud
- IBM
- Microsoft Azure
- NVIDIA
- Oracle cloud
- And many others
Final Thoughts
It seems like only yesterday I discussed the impact of Llama 3.1, and now Llama 3.2 is here to raise the bar even higher. The speed at which technology evolves is truly impressive. I'm excited to see how these advances will be applied across different industries and how they can positively influence our projects and solutions.
