Ling AI Model Family

Overview

The Ling AI Model Family, released by China's Ant Group (inclusionAI) in October 2025, represents a comprehensive suite of open-source AI models designed for all scales, all modalities, and complete open-source development. Rather than a single monolithic system, this initiative offers a family of specialized models across three main series:

• Ling (Large Language Models): Non-thinking models optimized for efficient language processing
• Ring (Thinking Models): Advanced reasoning models with explicit thinking capabilities
• Ming (Multimodal Models): Systems for processing text, images, audio, and video

The flagship model features 1 trillion total parameters with approximately 50 billion active parameters per token, representing a breakthrough in efficient large-scale AI. All models are released under MIT license, embodying the company's vision of AGI as a public good.

Model Family

Ling Series (Large Language Models)

The Ling series consists of non-thinking models optimized for efficient language processing:

• Ling-1T: Flagship trillion-parameter model with 50B active parameters, trained on 20+ trillion reasoning-dense tokens
• Ling-flash-2.0: Fast, efficient variant for high-throughput applications
• Ling-mini-2.0: Compact model balancing performance and resource efficiency
• Ling-lite-1.5: Lightweight model for resource-constrained environments
• Ling-lite: Ultra-efficient model for edge deployment
• Ling-plus: Enhanced variant with additional capabilities

Ring Series (Thinking Models)

The Ring series features advanced reasoning capabilities with explicit thinking mode:

• Ring-1T: World's first open-source trillion-parameter reasoning model (released September 2025)
• Ring-flash-2.0: Fast reasoning variant with efficient thinking pathways
• Ring-flash-linear-2.0: Linear-complexity reasoning for extended contexts
• Ring-mini-linear-2.0: Compact linear reasoning model
• Ring-mini-2.0: Small-scale thinking model with strong reasoning
• Ring-lite: Lightweight reasoning model for practical deployment

Ming Series (Multimodal Models)

The Ming series processes multiple modalities including text, images, audio, and video:

• Ming-flash-omni: Fast omnimodal model for all input types
• Ming-lite-omni: Efficient omnimodal model for resource-constrained settings
• Ming-UniAudio: Specialized audio understanding and generation
• Ming-UniVision: Advanced visual understanding and reasoning
• Ming-lite-uni: Lightweight unified multimodal model

Experimental Models

• LLaDA-MoE: Experimental architecture exploring novel MoE designs

Capabilities

The Ling AI Model Family provides comprehensive capabilities across all model series:

Language Processing (Ling Series)

• Flagship-Level Efficient Reasoning: Ling-1T extends the Pareto frontier of reasoning accuracy vs. length
• Advanced Code Generation: State-of-the-art performance in code generation and software development
• Aesthetic Understanding: Excels in visual reasoning and front-end code generation (#1 on ArtifactsBench)
• High Throughput: Flash variants optimized for rapid, efficient processing
• Resource Flexibility: Models from trillion-scale to edge deployment

Advanced Reasoning (Ring Series)

• Explicit Thinking Mode: Ring models feature advanced reasoning with visible thought processes
• World's First Open-Source Trillion-Parameter Reasoning Model: Ring-1T-preview (September 2025)
• Linear Complexity: Ring-linear variants for extended context reasoning
• Multi-Step Problem Solving: Strong performance on complex logical and mathematical tasks
• Efficient Reasoning Pathways: Optimized thinking modes across model sizes

Multimodal Processing (Ming Series)

• Omnimodal Understanding: Process text, images, audio, and video simultaneously
• Specialized Modalities: Dedicated audio (UniAudio) and vision (UniVision) models
• Cross-Modal Reasoning: Advanced understanding across different input types
• Flexible Deployment: From high-performance omni models to lightweight variants

Shared Capabilities

• Emergent Intelligence: Strong transfer capabilities at trillion-scale (approximately 70% tool-call accuracy with minimal training)
• Natural Language Understanding: Interprets complex instructions across all model types
• Cross-Platform Compatibility: Generates compatible code for multiple platforms
• Multilingual Support: Stylistically controlled content in multiple languages
• Open Source: All models available under MIT license

Ling-1T: Flagship Model

Ling-1T is the flagship non-thinking model featuring 1 trillion total parameters with approximately 50 billion active parameters per token. Pre-trained on over 20 trillion high-quality, reasoning-dense tokens, it demonstrates state-of-the-art performance on complex reasoning benchmarks while maintaining exceptional efficiency through its innovative MoE architecture and evolutionary chain-of-thought optimization.

Technical Specifications

Ling-1T incorporates cutting-edge architectural innovations:

• Model size: 1 trillion total parameters with approximately 50 billion active parameters per token (1/32 MoE activation ratio)
• Context window: 128K tokens (extendable with YaRN technique)
• Training data: 20+ trillion high-quality tokens with over 40% reasoning-dense data in later stages
• Architecture: Ling 2.0 architecture based on Mixture-of-Experts (MoE) Transformer design
• MoE Innovations:
  • MTP layers for enhanced compositional reasoning
  • Aux-loss-free sigmoid-scoring expert routing
  • Zero-mean updates for stable training
  • QK Normalization for fully stable convergence
• Training Precision: Largest known FP8-trained foundation model with 15%+ speedup and less than 0.1% loss deviation
• Optimization: Custom WSM (Warmup–Stable–Merge) LR scheduler with mid-train checkpoint merging
• Post-Training: Evolutionary Chain-of-Thought (Evo-CoT) for progressive reasoning enhancement
• Policy Optimization: LPO (Linguistics-Unit Policy Optimization) for sentence-level alignment

Use Cases

Ling-1T's efficient architecture makes it ideal for diverse applications:

• Software Development: High-quality code generation, debugging, and software architecture design
• Front-End Design: Creating aesthetically pleasing, functional user interfaces with visual reasoning
• Visual Reasoning: Complex visual understanding and design system implementation
• Mathematical Problem-Solving: Competition-level mathematics and complex calculations
• Tool Usage & Agents: Building AI agents with strong tool-call capabilities (70% accuracy on BFCL V3)
• Marketing & Content: Generating stylistically controlled, aesthetically refined marketing materials
• Cross-Platform Development: Creating compatible code for multiple platforms and frameworks
• Research & Analysis: Efficient processing of complex reasoning tasks with optimal token usage

Performance Metrics

Ling-1T demonstrates exceptional performance across multiple benchmarks:

• AIME 2025: Extends Pareto frontier of reasoning accuracy vs. reasoning length, showcasing efficient thinking
• ArtifactsBench: #1 ranking among open-source models for front-end generation and aesthetic understanding
• BFCL V3: Approximately 70% tool-call accuracy with minimal instruction tuning (no large-scale trajectory data)
• Code Generation: State-of-the-art performance in software development tasks
• Knowledge Benchmarks: Comprehensive evaluation across knowledge, code, math, reasoning, agent, and alignment tasks
• Efficiency: 15%+ end-to-end speedup with FP8 training, 40%+ utilization improvement with heterogeneous pipeline

Limitations

While the Ling AI Model Family has made strong progress, several considerations apply:

Ling Series Limitations

• Attention Mechanism: GQA-based attention is stable for long-context but relatively costly; future versions will adopt hybrid attention
• Agentic Capabilities: Room to grow in multi-turn interaction, long-term memory, and advanced tool use
• Instruction Alignment: Occasional deviations or role confusion may occur; ongoing improvements in alignment
• Non-Thinking Model: Designed for efficient reasoning without explicit thinking mode; for deep reasoning chains, use Ring series

Model Selection Guidance

• For explicit reasoning: Use Ring series (thinking models) instead of Ling series
• For multimodal tasks: Use Ming series for text, image, audio, or video processing
• For resource constraints: Consider lite variants across all series
• For maximum performance: Flagship models (Ling-1T, Ring-1T) require substantial compute resources

Safety & Alignment

Ling-1T incorporates comprehensive safety measures:

• Open Source Safety: Full transparency through MIT license and open model weights
• Alignment Optimization: LPO (Linguistics-Unit Policy Optimization) for precise reward-behavior alignment
• Training Stability: Superior training stability and generalization across reasoning tasks
• Evo-CoT Safety: Progressive reasoning enhancement under controllable cost constraints
• Community Oversight: Open-source model enabling community review and safety improvements

Innovation Highlights

Ling Scaling Law

The Ling 2.0 architecture was designed from the ground up using the Ling Scaling Law (arXiv:2507.17702), ensuring architectural and hyperparameter scalability even under 1e25–1e26 FLOPs of compute.

FP8 Mixed-Precision Training

Ling-1T is the largest known FP8-trained foundation model, achieving:

• 15%+ end-to-end speedup
• Improved memory efficiency
• Less than or equal to 0.1% loss deviation from BF16 across 1T tokens

Evolutionary Chain-of-Thought (Evo-CoT)

Built upon mid-training reasoning activation, Evo-CoT provides:

• Progressive reasoning enhancement under controllable cost
• Continual expansion of Pareto frontier (accuracy vs. efficiency)
• Ideal optimization for reflexive non-thinking models

Hybrid Reward Mechanism

The Syntax–Function–Aesthetics reward system enables:

• Correct and functional code generation
• Refined visual aesthetic understanding
• Superior performance on front-end generation tasks

Pricing & Access

The Ling AI Model Family offers flexible deployment options across all series:

API Access

• Model Access: Models available through various third-party API providers
• Format: OpenAI-compatible API format supported
• Self-Hosting: Full support for local deployment with vLLM and SGLang

Open Source

• License: MIT License - free for research and commercial use across all models
• Hugging Face: inclusionAI models with all model variants
• ModelScope: inclusionAI on ModelScope for faster downloads in China
• Deployment: Supports vLLM, SGLang, and other inference frameworks

Model Variants

• Ling Series: Access via inclusionai/ling-* model names
• Ring Series: Access via inclusionai/ring-* model names
• Ming Series: Access via inclusionai/ming-* model names

Ecosystem & Tools

The Ling AI Model Family is well-supported across deployment platforms:

• Hugging Face: Primary hub for all model weights and documentation
• ModelScope: Chinese platform for faster downloads
• vLLM: High-performance inference engine support for all series
• SGLang: Structured generation language support
• Custom Deployment: Full support for local and private cloud deployment across all models

Community & Resources

• Ant Group Press Release - Official announcement
• inclusionAI on Hugging Face - Complete model family and downloads
• Ling-1T on Hugging Face - Flagship model page
• Ling Scaling Law Paper - Architecture design principles
• WSM Scheduler Paper - Training methodology