Meta's newest open-source AI models compete with GPT, Gemini, and Claude.

      Meta has revealed the newest version of its open-source AI model family, Llama 4, as competition in the generative AI sector intensifies. The updated AI family comprises four models, including Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to details shared on its AI website, the models were trained on “extensive amounts of unlabeled text, images, and video data,” suggesting they will possess diverse multimodal capabilities.

      Currently, users can access two of the models — Llama 4 Scout and Llama 4 Maverick — on Meta’s platforms, including WhatsApp, Messenger, and Instagram Direct, as well as on Meta’s AI website, Llama.com, starting Saturday. Developers can also find AI models in open-source repositories like Hugging Face. The Llama 4 Behemoth model has not yet been released and is still undergoing training. The company has indicated that this model is expected to outperform its peers and will act as an educator for the other Llama 4 models.

      During its internal testing, Meta compared the Llama 4 models with competitor AI technologies to establish their capabilities and ideal applications. The company stated that Llama 4 Maverick is particularly well-suited for creative writing. Additionally, it surpassed the performance of OpenAI's GPT-4o and Google’s Gemini 2.0 in areas such as coding, reasoning, multilingual tasks, long-context handling, and image generation. However, Maverick struggled against models like Gemini 2.5 Pro, GPT-4.5, and Anthropic Claude 3.7 Sonnet.

      While Meta asserts that Behemoth can outperform most competing models, with the exception of Gemini 2.5 Pro, it has struggled to bring down the hardware costs related to training its most advanced model. TechCrunch noted that the attention garnered by the Chinese AI firm DeepSeek, with its competitive and cost-effective models, has caught Meta's attention. The company has reportedly been closely analyzing how DeepSeek developed its prominent models, including R1 and V3, at lower operational costs compared to prior Llama versions.

      Meta explained that the Llama 4 Scout model can operate on a single Nvidia H100 GPU, while the Llama 4 Maverick model requires one Nvidia H100 DGX graphics system. Meta is scheduled to host its inaugural LlamaCon AI conference on April 29. The company also plans to launch a standalone Meta AI chatbot in the second quarter of the year, as reported by CNBC.

      Meta is not the only company being cautious with the rollout of its major AI models. OpenAI recently postponed the release of its GPT-5 model. CEO Sam Altman announced on social media that enthusiasts can expect new reasoning models, o3 and o4-mini, in the coming weeks as alternatives to GPT-5. He specified that GPT-5 is now set to launch in the forthcoming months, allowing OpenAI extra time to refine the model.