Ling-1T

Ling-1T is positioned as the first flagship non-thinking model in the Ling 2.0 series, purpose-built for efficient reasoning and scalable cognition. It is a Mixture-of-Experts style model that pairs roughly one trillion total parameters with about 50 billion active parameters per token, a configuration designed to keep inference cost manageable while preserving the capacity of a trillion-parameter network. The base model is pre-trained on more than 20 trillion high-quality, reasoning-dense tokens and supports the cataloged API limit context window, making it well suited to long documents, multi-step code, and competition-style problem solving where the model can keep extensive chains of thought in working memory.

Beyond raw scale, Ling-1T is shaped by an evolutionary chain-of-thought curriculum, called Evo-CoT, that runs through both mid-training and post-training to deepen reasoning while controlling verbosity. In evaluations against leading open-source and closed-source systems on code generation, software development, competition mathematics, professional mathematics, and logical reasoning, the Ling-1T card reports consistently strong complex-reasoning results, with the model extending the Pareto frontier of accuracy versus reasoning length on benchmarks such as AIME 25. Distributed as open weights under the inclusionAI organization on Hugging Face and ModelScope, Ling-1T is a practical fit for teams that need a self-hostable, long-context reasoning engine for math, code, and analytical workloads.