inclusionAI/Ling-1T

5.0 based on 1 review
Warm
Public
1000B
FP8
32768
License: mit
Hugging Face
Overview

Ling-1T: A Trillion-Parameter Model for Efficient Reasoning

Ling-1T, developed by inclusionAI, is the flagship non-thinking model in the Ling 2.0 series, featuring 1 trillion total parameters with approximately 50 billion active parameters per token. It is pre-trained on over 20 trillion high-quality, reasoning-dense tokens and supports an extended context length of up to 128K.

Key Capabilities

  • Flagship-Level Efficient Reasoning: Consistently demonstrates superior complex reasoning across code generation, software development, mathematics, and logical reasoning, often outperforming both open-source and closed-source models like GPT-5-main and Gemini-2.5-Pro.
  • Evolutionary Chain-of-Thought (Evo-CoT): Employs Evo-CoT during mid-training and post-training to enhance reasoning efficiency and depth, achieving a strong balance between accuracy and efficiency.
  • Aesthetic Understanding & Front-End Generation: Excels in visual reasoning and front-end code generation, utilizing a hybrid Syntax–Function–Aesthetics reward mechanism. It ranks first among open-source models on ArtifactsBench.
  • Emergent Intelligence: Exhibits strong emergent reasoning and transfer capabilities at the trillion-parameter scale, including complex natural language interpretation, functional visual component transformation, and cross-platform front-end code generation.
  • Advanced Architecture: Built on the Ling 2.0 architecture with innovations like 1/32 MoE activation ratio, MTP layers, Aux-loss-free sigmoid-scoring expert routing, and QK Normalization for stable convergence.
  • FP8 Training: The largest FP8-trained foundation model to date, achieving over 15% end-to-end speedup with minimal loss deviation.

Good For

  • Applications requiring highly efficient and precise complex reasoning.
  • Tasks involving code generation, software development, and advanced mathematics.
  • Use cases demanding visual aesthetic understanding and front-end code synthesis.
  • Scenarios benefiting from long context processing (up to 128K tokens).

Ling-1T aims to advance general, collaborative human–AI intelligence, offering a powerful tool for developers seeking high-performance, reasoning-focused models.