Ling-2.6-flash

Ling-2.6-flash is an instruction-tuned model from Ant Group's inclusionAI team, part of the Ling family. It uses a highly sparse 104 billion parameter Mixture-of-Experts architecture that activates only 7.4 billion parameters per token, so it draws on a large parameter pool while keeping inference cost close to that of a much smaller dense model. The model is tuned specifically for agentic workflows, with strengths in tool calling, multi-step planning, and structured function dispatch. It was trained with agentic reinforcement learning to handle terminal operations and sustained chains of tool calls while keeping token consumption low. With a 262K token context window, it suits long-document processing, codebase-scale reasoning, and multi-step agent traces. It supports text input, tool use, and structured output.