ERNIE-4.5-21B-A3B-Thinking

ERNIE-4.5-21B-A3B-Thinking is Baidu's lightweight mixture-of-experts model optimized for deep reasoning and multi-step problem-solving despite activating only 3 billion parameters. It targets logical reasoning, mathematics, science, coding, and academic benchmarks that typically require far larger dense parameter counts. The 131K token context window enables extended document analysis, academic paper integration, and multi-file codebase processing. Built on the same heterogeneous MoE architecture as the base 21B model but enhanced with explicit thinking for complex tasks, it is released as open weights under the Apache 2.0 license, supports multiple inference frameworks (FastDeploy, vLLM, Transformers), and offers quantized builds for resource-constrained environments.