Meta has released the latest version of its Llama family of artificial intelligence models.

According to the company, this is a new upgrade of the entire family, called Llama 3.1, and includes inferential improvements to the small and medium versions.

They also remain open source, meaning that any company or organization can download, tweak, and run the models on their own hardware.

According to Meta, Llama 3.1 405b is the largest of the family and is comparable to OpenAI and Google models. This includes inference and coding capabilities.

It is available on the Groq Cloud inference platform, Meta's own MetaAI chatbot, and developer clouds like AWS, IBM, and Azure.

Llama 3.1 405b (meaning 405 billion parameters) is a major change for both Meta and the open source AI community, and the company claims to surpass Claude 3.5 Sonnet and GPT-4o in many key benchmarks.

While Meta will use this for its MetaAI chatbots, the model brings advanced frontier-grade intelligence to companies, allowing them to install it on their own hardware, adapt it to their own needs, and pay OpenAI a per token fee They can use it without having to pay a per-token fee for OpenAI.

During training and subsequent fine-tuning, Meta says it focused on the “helpfulness” of the chatbot. In a statement, they explain: “With Llama 3.1 405B, we worked to improve the model's helpfulness, quality, and detailed instruction-following ability in response to user instructions, while ensuring a high level of security.

A context window of 128,000 tokens was included, bringing it nearly on par with industry standards.”

Llama 3.1 comes in three sizes. The first is 8B, which is lightweight, super fast, and can run anywhere, including smartphones. [It runs primarily in data centers, but also runs on good gaming PCs.

Finally, there is the new flagship foundation model 405B, which approaches the same quality and features as leading private models like OpenAI, Google, and Anthropic.

Llama 3.1, which includes the new 405b version, is available on the MetaAI chatbot, but has a limited worldwide release. Wider coverage can be accessed using Groq Cloud or the very quick Groq chatbot interface. [Microsoft, Google, Amazon, DataBricks, and other developer platforms have also loaded the new model, and if you have your own data center, you can download the three versions from GitHub and run them yourself.

Two smaller versions, 8b and 70B, are available on Ollama, a simple installer for local AI models on Windows, Mac, and Linux.

Meta CEO Mark Zuckerberg predicts that Meta AI will become the most widely used assistant by the end of this year, surpassing ChatGPT as the #1 way to access AI.

He says this is thanks to expanded access to assistants through WhatsApp, Instagram, Ray-Ban smart glasses, and standalone MetaAI chatbots. It will also be included in Quest's mixed reality headset to replace current voice commands.