With the release of Llama 3.1, Meta has declared its intention to compete directly with proprietary models such as OpenAI’s GPT-4o and Google’s Gemini 1.5.
Llama 3.1 is the first open source “frontier model” that, despite the limitations of its license, promises great customization capacity for companies and developers, positioning Meta’s AI proposal as a leader in the industry.
Llama 3.1 is available in various sizes, and the largest model, “405B,” has 405 billion neural parameters, surpassing other leading open source models such as Mixtral and Google Gemma 2.
Being open source, anyone can train and run Llama 3.1 on their own computers, as long as they are extremely powerful. Let’s explore together how this new Meta model can transform the current AI landscape.
The trailer for Llama 3.1
The models of the Llama 3.1 family stand out for their size and exceptional capabilities. The flagship model, with 405 billion parameters, is designed to push the limits of AI performance and usability.
Mark Zuckerberg, CEO of Meta, has expressed in a letter his belief that “open” models are the future. Therefore, the Meta con Llama 3.1 approach emphasizes openness, unlike “closed” models such as GPT-4, Gemini 1.5 or Claude 3.5 Sonnet.
Although Meta describes Llama 3.1 as open source, some experts debate this claim. According to Bradley Shimmin, industry analyst at Omdia, Llama models are not truly open source as defined by the Open Source Initiative.
A more precise term would be to say that Llama 3.1 has open “weights” or parameters. This means that these models can be tuned and trained, although Meta does not offer complete transparency into the data used to pre-train the models.
The release also includes smaller variants, such as 70B and 8B, that suit different use cases and computing resources.
What do I need to run Llama 3.1?
Running the largest model, Llama 3.1-405B, requires an extremely powerful computing infrastructure. Specifically, you would need at least two Nvidia H100 servers, which combine a total of 16 GPUs, 1 TB of RAM and 870 GB of VRAM.
To use Llama 3.1-405B in practical applications, it makes most sense to use cloud infrastructure provided by services such as AWS, Microsoft Azure or Nvidia DGX Cloud. This facilitates its implementation without having to have the necessary hardware.
The smaller, less powerful versions of Llama 3.1 have more modest hardware needs. For example, Llama 3.1 8B would “only” need about 16 GB of VRAM, making it possible to run it on a system with an Nvidia 4090.
The annual cost to run the Llama 3.1-405B model on suitable infrastructure can exceed €300,000, making it accessible primarily to large companies with significant AI research and development budgets.
To mitigate the high costs of running the 405B model, companies are expected to turn to cloud services.
Meta has established partnerships with a wide range of companies, including AWS, Google Cloud, Microsoft Azure, and IBM watsonx, to provide Llama 3.1 models via APIs. This approach allows companies to access the necessary IT resources based on their needs, reducing initial investments.
Key advantages of Llama 3.1 for businesses
One of the most significant advantages of Llama 3.1 for businesses is the ability to customize models for specific use cases without incurring high costs.
According to Paul Nashawaty, principal analyst at The Futurum Group, providing a language model with open weights allows companies to build custom AI solutions without having to pay expensive licenses.
Performance comparable to more advanced models
The Llama 3.1 family, particularly the 405B model, has demonstrated competitive performance in benchmark tests against leading proprietary models such as OpenAI’s GPT-4o and Google’s Gemini 1.5.
Meta’s blog post highlights that the 405B model outperformed or equaled these models on tests such as MMLU, MATH, GSM8K, and ARC Challenge, which assess general intelligence, mathematics, and reasoning abilities.
Data security and privacy
Arnal Dayaratna, vice president of research at IDC, notes that companies can tune Llama 3.1 models using their own data without having to share data with a third-party vendor.
This capability helps avoid an increasingly common problem with proprietary LLMs; become dependent and trapped by a single AI provider.
How will the competition respond to Llama 3.1?
The release of Llama 3.1 represents a significant challenge for proprietary LLM providers. Analysts predict that the openness and high performance of Meta models could disrupt the market.
Tobias Zwingmann, managing partner at Rapyd.AI, suggests that the availability of open models like Llama 3.1 could lead companies to reduce their reliance on closed proprietary LLMs.
This change could disrupt companies and suppliers that build and sell proprietary models.
The competition is already responding to Meta’s move. Within 24 hours of launching Llama 3.1, OpenAI announced a free tier to customize its GPT-4o mini model.
This reaction indicates a broader trend toward reducing costs and increasing accessibility to stay competitive.
Is Llama 3.1 available in the European Union?
Llama 3.1 is available globally, without geographical restrictions, but can only be used through a Cloud Computing provider or running it locally. However, it is not yet available from the Meta.AI multimodal interface in the European Union and other countries.
Meta perceives the regulatory environment in the EU as unpredictable, which adds a layer of uncertainty when implementing new technologies.
For this reason, the parent company of Facebook and Instagram has decided not to launch its multimodal model Llama 3.1 in the European bloc due to the strict and sometimes unpredictable GDPR (General Data Protection Regulation) regulations.
The future of open LLMs
Flame Meta 3.1 represents a significant milestone in the development of large language models. Its openness, performance and flexibility offer substantial benefits to businesses, while challenging the dominance of proprietary model vendors.
As the AI landscape evolves, the release of Llama 3.1 is expected to drive greater competition, innovation, and adoption of open AI models.
For businesses, the ability to leverage powerful, customizable AI solutions without incurring high costs marks a new era of possibilities in artificial intelligence.
This post is also available in: