Meta has recently unveiled the newest version of its open-source AI model series, known as Llama 4. This development comes at a time when competition in the generative AI industry is growing more intense.
The latest AI family consists of four models, namely Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. Meta has disclosed that these models were trained on a vast amount of unlabeled text, image, and video data, suggesting that they possess diverse multimodal capabilities.
Currently, two models – Llama 4 Scout and Llama 4 Maverick – are accessible to users on Meta’s platforms such as WhatsApp, Messenger, and Instagram Direct, as well as on Meta’s AI website, Llama.com. Developers can also access these AI models on open-source repositories like Hugging Face. The third model, Llama 4 Behemoth, is still undergoing training and has not been released yet. Meta has hinted that Behemoth is expected to surpass other models in performance and will act as a mentor to the rest of the Llama 4 family.
During internal testing, Meta compared the Llama 4 models to competitor AI technologies to assess their capabilities and ideal applications. Llama 4 Maverick was found to excel in creative writing, outperforming models like OpenAI GPT-4o and Google Gemini 2.0 in coding, reasoning, multilingual tasks, long-context scenarios, and image generation. Despite this, Maverick fell short compared to models like Gemini 2.5 Pro, GPT-4.5, and Anthropic Claude 3.7 Sonnet.
Although Meta claims that Behemoth can outperform most models, except for Gemini 2.5 Pro, the company has yet to address the high hardware costs associated with training this powerful model.
DailyTech reported that the rise of the Chinese AI company DeepSeek, known for its cost-effective models like R1 and V3, has prompted Meta to rethink its strategy. Meta has been closely studying DeepSeek’s approach to developing high-performance models at lower costs compared to previous Llama models.
Meta revealed that the Llama 4 Scout model can operate on a single Nvidia H100 GPU, while the Llama 4 Maverick model requires a Nvidia H100 DGX graphics system.
Meta is gearing up to host its inaugural LlamaCon AI conference on April 29. Additionally, CNBC has reported that a standalone Meta AI chatbot is scheduled for release in the second quarter of the year.
OpenAI, another prominent player in the AI industry, recently postponed the launch of its GPT-5 model. Instead, the company’s CEO, Sam Altman, announced the upcoming release of o3 and o4-mini reasoning models as an alternative to GPT-5. Altman mentioned that GPT-5 will be delayed for a few months to ensure it meets the required standards.
