Meta releases Llama 3 open models

Meta has released its next-generation open source LLM, Llama 3, and made it available via Hugging Face. The models consist of an 8B model for efficient deployment and a 70B model for large-scale applications, each provided in Base and Instruct-tuned versions. Llama Guard 2 was also released for safe model usage.

The key technical changes are as follows.

• Expanded tokenizer: The vocabulary size was significantly increased from the previous 32K to 128,256, improving text encoding efficiency and multilingual performance.
• GQA adoption: Grouped-Query Attention (GQA) was applied to the 8B model to improve long-context processing efficiency.
• Large-scale training: The models were trained on approximately 15 trillion tokens, 8 times more than the previous generation.

Hugging Face supports an ecosystem for immediately using Llama 3 through Transformers integration, Hugging Chat support, Inference Endpoints, and integration with major clouds (Google Cloud, AWS).