Meta releases Llama 3 open source model

Meta has released the latest model in its Llama series of open generative AI models, Llama 3. Llama 3 8B contains 8 billion parameters, and Llama 3 70B contains 70 billion parameters.

The Llama 3 8B beats other open models such as the Mistral 7B and Google Gemma 7B in nine tests. Meta* has also developed its own set of assignments, covering a variety of use cases, from programming and creative writing to reasoning and summarizing. Llama 3 70B took first place among Mistral's Mistral Medium, OpenAI's GPT-3.5 and Claude Sonnet.

In terms of quality, the company notes that users of the new Llama models should expect greater controllability, less likelihood of failure, and greater accuracy in simple matters. This was achieved thanks to a large data set: a collection of 15 trillion tokens or approximately 750 billion words.

In addition, Meta* says it has developed new data filtering pipelines to improve the quality of training information and updated a pair of generative AI security suites, Llama Guard and CybersecEval. The company is also releasing a new tool, Code Shield, designed to detect code that may contain security vulnerabilities.

Llama 3 8B and Llama 3 70B will be integrated into the Meta AI virtual assistant, which the company is positioning as the most sophisticated of its free options. Models are available for download now. However, despite Meta*'s claims that its Llama family of models is open source, the company prohibits developers from using its models to train other generative systems.

Meta also states that it is currently training the Llama 3 model with over 400 million parameters. The company notes, “Our goal in the near future is to make Llama 3 multilingual and multimodal, create longer context, and improve the overall performance of core capabilities such as reasoning and coding.”