Select Language:
Meta has unveiled its newest version of the open-source AI family, known as Llama 4, amidst growing competition in the generative AI sector.
This latest lineup comprises four models, specifically Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to information shared on Meta’s AI website, these models were trained using extensive datasets of unlabeled text, images, and videos, showcasing their diverse multimodal capabilities.
As of this past Saturday, the Llama 4 Scout and Llama 4 Maverick models are accessible to users across various Meta platforms, including WhatsApp, Messenger, and Instagram Direct, as well as on Meta’s dedicated AI site, Llama.com. Developers can also find these AI models available in open-source repositories like Hugging Face. The Llama 4 Behemoth model, however, remains in training and has not yet been released. Meta has indicated that the Behemoth model is expected to surpass its counterparts and serve a pivotal role in guiding the other models within the Llama 4 series.
While internally testing the Llama 4 models, Meta conducted comparisons against competitor AI technologies to assess their capabilities and ideal applications. The company highlighted that Llama 4 Maverick excels in creative writing, outperforming models such as OpenAI’s GPT-4o and Google’s Gemini 2.0 in areas like coding, reasoning, multilingual comprehension, long-context processing, and image generation. However, Maverick faced challenges matching the performance of newer models like Gemini 2.5 Pro, GPT-4.5, and Anthropic Claude 3.7 Sonnet.
Despite Meta’s assertions that the Behemoth model will outstrip most models—including Gemini 2.5 Pro—the company is still facing challenges in minimizing the hardware costs associated with training its most formidable model.
TechCrunch has observed that the Chinese AI firm DeepSeek has gained significant traction with its competitively priced models, prompting Meta to closely examine how the rival company managed to develop impactful models like R1 and V3 at lower operational costs than previous iterations of Llama.
Notably, the Llama 4 Scout model is capable of operating on a single Nvidia H100 GPU, while the Llama 4 Maverick model requires a Nvidia H100 DGX graphics system to function.
Meta plans to host its inaugural LlamaCon AI conference on April 29. The company also intends to launch a standalone Meta AI chatbot within the second quarter of the year.
In a parallel development, OpenAI has adjusted its GPT-5 model timeline, with CEO Sam Altman announcing on social media that users should anticipate new reasoning models (o3 and o4-mini) in the upcoming weeks as alternatives to GPT-5. Altman confirmed that GPT-5 will be released in the upcoming months, allowing OpenAI more time to refine the model.
ChatGPT
Perplexity
Gemini AI
Grok AI
