ERNIE 4.5 21B A3B

ERNIE 4.5 21B A3B is Baidu's efficient text-only mixture-of-experts model, activating just 3 billion parameters per token despite 21 billion total parameters. Built on a heterogeneous MoE architecture with modality-isolated routing, it handles general-purpose English and Chinese understanding, generation, and instruction following with native support for function calling. Designed for both research and production use, the model targets solid reasoning and comprehension across STEM and knowledge tasks. It supports a 120K token context for processing lengthy documents and ships as open weights under the Apache 2.0 license. Quantized variants (4-bit, 2-bit) are available for more efficient deployment.