: Designed for efficiency, this model has 17 billion active parameters. It fits on a single H100 GPU. It is optimized for high-speed performance (up to 460+ tokens per second) and long-document reasoning.
: The models use a "mixture of experts," where only a subset of the total parameters (e.g., 17 billion active parameters in the Scout model) are activated for any given task. This significantly reduces computational costs and latency while maintaining high performance. Laskamp4
: This is a larger model with 400 billion parameters and 128 experts. It rivals top proprietary systems like GPT-4 and Gemini in complex reasoning, coding, and image understanding. : Designed for efficiency, this model has 17
: Previews suggest this is Meta's most powerful model yet. It serves as a "teacher" for smaller models through distillation processes. Reception and Performance : The models use a "mixture of experts,"
The Llama 4 series represents a major shift in open-source artificial intelligence, moving toward capabilities and Mixture-of-Experts (MoE) architectures.
: Unlike previous versions that relied on "bolted-on" vision components, Llama 4 was trained from the start with text, images, and video frames.