Meta’s introduction of two Llama 4 artificial intelligence (AI) models has established the U.S. as a frontrunner in the AI competition, according to David Sacks, the nation’s AI and crypto czar. He shared this insight in a post on X over the weekend, stating:
“To secure victory in the AI race, the U.S. must also triumph in open source, and Llama 4 brings us back to the forefront.”
On April 6, Meta revealed its fourth-generation open-source Llama 4 models, known as Llama 4 Scout and Llama 4 Maverick.
The AI competition escalates with DeepSeek’s debut
DeepSeek, a Chinese AI startup established in 2023, introduced its inaugural model in December 2024. By January 2025, it launched a chatbot that claimed to compete with the capabilities of OpenAI’s ChatGPT.
The downloads of DeepSeek’s generative AI, called DeepSeek R1, skyrocketed on app stores, challenging the prevailing notion that substantial investments and costly chips are essential for success in the AI landscape.
This development led to declines in stock prices for major U.S. tech firms like Nvidia following the launch of DeepSeek R1.
DeepSeek asserted that they invested only about $6 million in training their AI model, in stark contrast to the reportedly $100 million that OpenAI allocated to train ChatGPT-4.
Venture capitalist Marc Andreessen characterized DeepSeek’s R1 launch as “AI’s Sputnik moment,” while U.S. President Donald J. Trump referred to it as a “wake-up call” for American companies.
In the wake of these developments, U.S. companies have been striving to make significant advancements in AI, with Sacks emphasizing that Llama 4 has been pivotal in those efforts.
Meta asserts Llama 4 models are “top of the line”
Meta claims that Llama 4 represents their “most sophisticated models to date” and stands as “the best in its class for multimodality.” Both models are currently accessible for download and integration on Meta’s platforms like WhatsApp and Instagram.
Multimodal AI systems can handle diverse data types—such as text, images, audio, and video—simultaneously. This capability allows AI to understand intricate scenarios and deliver more thorough responses.
Llama 4 Scout and Llama 4 Maverick are the first open-source Meta AI models utilizing a mixture of experts (MoE) architecture. In an MoE framework, multiple smaller models or specialized experts cooperate to enhance the performance of the larger AI model. This setup allows experts to tackle the specific areas they are optimized for.
Llama 4 Scout features 17 billion active parameters and employs 16 experts. Conversely, Llama 4 Maverick also has 17 billion parameters but integrates 128 experts. The former can operate on a single NVIDIA H100 GPU, while the latter necessitates an H100 host.
Meta claims that Llama 4 Scout surpasses Gemma 3, Gemini 2.0 Flash-Lite, and Mistral 3.1 across a series of well-known benchmarks.
On the other hand, Llama 4 Maverick yields performance akin to DeepSeek v3 in reasoning and coding, despite having fewer than half the active parameters. Additionally, Meta contends that Llama 4 Maverick outperforms GPT-4o and Gemini 2.0 Flash in various benchmarks.
Moreover, according to Meta’s assessments, Llama 4 “exhibits a strong political leaning at a rate similar to Grok.”
Additionally, Meta has introduced Llama 4 Behemoth, which is still undergoing training and is touted as one of the “smartest” large language models (LLMs) globally.
The first Llama model was launched by Meta in February 2023.
Related Content
