InclusionAI Releases Ling-3.1-flash with 560B Parameters
InclusionAI has released Ling-3.1-flash, a massive hybrid reasoning Mixture-of-Experts model featuring 560 billion total parameters and 25 billion active parameters per token. This represents a massive scale-up from their previous version 3.0, which only had 127 billion parameters. The release introduces a high-capacity model architecture that drastically increases total parameters while keeping active parameters low for efficient inference. This highlights the ongoing trend of leveraging advanced Mixture-of-Experts designs to balance high model intelligence with manageable operational speeds. While the Hugging Face listing notes a 1-million-token context window, provider implementations such as OpenRouter report support for a 262,144-token context window with up to 32,768 completion tokens. The model functions as a hybrid reasoning chat system accessible via API and platform routing.
## BACKGROUND
Mixture-of-Experts is a machine learning architecture where a model divides its capacity into multiple specialized subnetworks called experts, routing each incoming token through only a small subset of them. This allows the model to maintain a massive total parameter count for vast knowledge while requiring the computing power of a much smaller model per token.