--- source: 'https://howaiworks.ai/blog/ant-group-ling-1t-announcement' section: blog title: 'Ant Group Unveils Ling-1T: Trillion-Param Model' description: >- Ant Group releases Ling-1T, a trillion-parameter open-source AI model with state-of-the-art coding, reasoning, and multimodal capabilities. date: '2025-10-13' author: HowAIWorks Team tags: - AI - ant-group - ling-1t - ai-models - open-source - Coding - Reasoning - Multimodal - artificial-intelligence - china-ai - trillion-parameter-model - chinese-ai - moe-architecture - fp8-training - open-source-llm readingTime: 7 minutes newsSource: Ant Group breakingNews: false imageWidth: 1200 imageHeight: 630 --- # Ant Group Unveils Ling-1T: Trillion-Param Model > Ant Group releases Ling-1T, a trillion-parameter open-source AI model with state-of-the-art coding, reasoning, and multimodal capabilities. ## Introduction The Chinese fintech giant **Ant Group** has unveiled the **[Ling AI model family](https://howaiworks.ai/models/ling)** and launched **Ling-1T**, a trillion-parameter language model that sets new standards in code generation, mathematical reasoning, and logical problem-solving. This comprehensive release introduces three distinct model series designed to address different AI capabilities while maintaining the company's commitment to open-source development and accessible artificial general intelligence (AGI). ## Quick Facts - **Release Date**: October 13, 2025 - **Developer**: Ant Group (inclusionAI) - **Model Scale**: 1 trillion total parameters, ~50B active per token - **License**: MIT (Free & Open Source) - **Top Performance**: 70.42% on AIME 2025, #1 on ArtifactsBench (open-source) - **Availability**: [Hugging Face](https://huggingface.co/inclusionAI) & [ModelScope](https://modelscope.cn/organization/inclusionAI) ## Ling AI Model Family Overview ### Three Distinct Series Ant Group's new Ling family (also known as BaiLing) comprises three specialized series: **Ling Series**: - **Focus**: Efficient language processing with implicit reasoning (non-thinking models) - **Use cases**: Content generation, code development, language translation, text processing - **Architecture**: Optimized for efficient reasoning without explicit thinking mode - **Flagship model**: Ling-1T with one trillion parameters **Ring Series**: - **Focus**: Advanced reasoning capabilities - **Breakthrough**: World's first open-source trillion-parameter reasoning model (Ring-1T-preview) - **Use cases**: Complex problem-solving, logical reasoning, mathematical computations - **Previous release**: Ring-1T-preview launched in September **Ming Series**: - **Focus**: Multimodal processing capabilities - **Capabilities**: Processing and understanding multiple data types (text, images, audio, video) - **Use cases**: Cross-modal understanding, multimedia analysis, integrated AI applications - **Integration**: Combines language and visual understanding ## Key Highlights of the Release ### Breakthrough Performance Ling-1T sets new standards with verified benchmarks: - **AIME 2025**: Achieved **70.42% accuracy** on competitive mathematics, using 4,000+ output tokens per problem—matching industry-leading models - **ArtifactsBench**: Ranks **#1 among open-source models** for front-end generation and aesthetic understanding - **BFCL V3**: Demonstrates **approximately 70% tool-call accuracy** with minimal instruction tuning, showing strong emergent capabilities - **Efficient Architecture**: 1T total parameters but only approximately 50B active per token (1/32 MoE ratio), achieving 15%+ speedup with FP8 training ### Technical Innovations The release introduces several architectural breakthroughs: - **Ling Scaling Law**: Purpose-built architecture for trillion-scale efficiency under 1e25–1e26 FLOPs - **Evo-CoT Optimization**: Evolutionary Chain-of-Thought for progressive reasoning enhancement - **LPO Training**: Linguistics-Unit Policy Optimization for precise sentence-level alignment - **FP8 Training**: Largest known FP8-trained foundation model with less than 0.1% loss deviation ## What Makes This Release Significant ### First Complete Open-Source AI Family at Trillion Scale This release marks a milestone in open-source AI development: - **Unprecedented Scale**: First complete family of trillion-parameter models (Ling, Ring, Ming) all open-sourced under MIT license - **No Cost Barriers**: Eliminates licensing costs that typically run into millions for comparable proprietary models - **Production-Ready**: All models deployed and available via API, with comprehensive deployment support (vLLM, SGLang) - **World First**: Ring-1T-preview was the world's first open-source trillion-parameter reasoning model (September 2025) ## Open Source Commitment ### Community Access Ant Group has made Ling-1T and Ring-1T-preview fully open source: - **Model weights**: Free access to complete model parameters - **Training data**: Transparency about training methodology - **Implementation details**: Comprehensive documentation for deployment - **Community support**: Active engagement with open-source community - **No restrictions**: Minimal barriers to commercial and research use ### AGI as a Public Good Ant Group's Chief Technology Officer, He Zhenyu, articulates the company's vision: > "We believe that artificial general intelligence (AGI) should be a public good—a shared achievement for humanity's intelligent future. We are committed to building practical and inclusive AGI services that benefit everyone, which requires continuous advancement of technology." ### Industry Impact The open-source release has significant implications: - **Research acceleration**: Enables widespread experimentation and innovation - **Cost reduction**: Eliminates licensing costs for trillion-parameter models - **Collaboration**: Facilitates global cooperation on AI advancement - **Accessibility**: Makes advanced AI capabilities available to smaller organizations - **Transparency**: Promotes understanding of large-scale AI systems ## Deployment and Accessibility [Ling-1T](https://howaiworks.ai/models/ling) and the entire model family are immediately available: - **Open Source**: Full model weights on [Hugging Face](https://huggingface.co/inclusionAI) and [ModelScope](https://modelscope.cn/organization/inclusionAI) under MIT license - **API Access**: Available through third-party providers with OpenAI-compatible API format - **Self-Hosting**: Full support for vLLM and SGLang deployment - **No Restrictions**: Free for commercial and research use ## Competitive Positioning ### Global AI Landscape Ling-1T positions Ant Group among AI leaders: - **Chinese AI innovation**: Demonstrates China's advancement in large language models - **Open-source leadership**: Sets example for open AI development - **Technical capability**: Proves competitiveness with Western AI models - **Scale achievement**: Joins elite group of trillion-parameter models - **Performance parity**: Matches or exceeds comparable international models ### Market Differentiation Key advantages of the Ling family: - **Specialized series**: Targeted models for different use cases - **Open access**: No cost barriers for researchers and developers - **Comprehensive capabilities**: Coverage across reasoning, language, and multimodal tasks - **Strong performance**: Competitive results on industry benchmarks - **Community-driven**: Open development model encouraging collaboration ## Impact on the AI Ecosystem ### Democratizing Advanced AI The release significantly lowers barriers to entry for advanced AI: - **Cost Savings**: Organizations can now deploy trillion-parameter models without licensing fees that typically cost millions - **Research Acceleration**: Academic institutions gain access to frontier-level AI for research and education - **Startup Enablement**: Smaller companies can compete with tech giants using the same AI infrastructure - **Global Access**: Models optimized for deployment in diverse environments, from cloud to edge devices ### Advancing Open-Source AI This release strengthens the open-source AI movement: - **Transparency**: Full access to model architecture, training methodology, and weights - **Reproducibility**: Complete documentation enables verification and improvement by the community - **Innovation**: Creates foundation for derivative works and specialized applications - **Collaboration**: Facilitates global cooperation on AI safety and capability research ## Industry Reaction and Context ### Market Impact The release comes at a critical time for the AI industry: - **Competitive Pressure**: Adds significant pressure on proprietary models from OpenAI, Anthropic, and Google to justify premium pricing - **Open-Source Leadership**: Positions China's AI ecosystem as a leader in open-source development, following DeepSeek's success - **Performance Parity**: Demonstrates that open-source models can match closed-source alternatives on key benchmarks - **Ecosystem Growth**: Expected to spawn numerous derivative applications and specialized fine-tunes ### Strategic Timing Ant Group's release aligns with broader industry shifts: - **Post-Scaling Laws Era**: Focus on efficiency and specialized architectures rather than pure parameter scaling - **AGI Democratization**: Growing movement toward AGI as a public good, not proprietary advantage - **Multi-Model Strategies**: Trend toward model families (Ling/Ring/Ming) rather than single flagship models - **Production Readiness**: Emphasis on deployment infrastructure alongside model release ## Conclusion Ant Group's release of the **[Ling AI model family](https://howaiworks.ai/models/ling)** and the trillion-parameter Ling-1T model represents a significant milestone in open-source artificial intelligence. By combining state-of-the-art performance with complete open access, Ant Group is advancing its vision of AGI as a public good while demonstrating competitive technical capabilities. ### Key Takeaways: - **Complete Model Family**: First open-source release of an entire trillion-parameter AI family (Ling, Ring, Ming) under MIT license - **Verified Performance**: 70.42% on AIME 2025, #1 on ArtifactsBench (open-source), approximately 70% on BFCL V3—matching proprietary models - **Technical Innovation**: Largest FP8-trained model, novel Evo-CoT optimization, and LPO alignment methodology - **Immediate Availability**: Production-ready deployment via Hugging Face with vLLM/SGLang support and third-party APIs - **World Firsts**: Ring-1T as first open-source trillion-parameter reasoning model; Ling-1T as most efficient at this scale - **Market Impact**: Disrupts pricing models of proprietary AI, enabling startups and researchers to access frontier-level capabilities - **AGI Philosophy**: Embodies Ant Group's vision of AGI as a public good, not proprietary advantage This release marks a **pivotal moment in AI democratization**, proving that open-source models can achieve parity with closed alternatives while eliminating cost barriers that have limited innovation to well-funded organizations. ## Sources - [Ant Group Press Release - Ling-1T Announcement](https://www.antgroup.com/en/news-media/press-releases/1759982400000) - [Ling-1T on Hugging Face](https://huggingface.co/inclusionAI/Ling-1T) - [inclusionAI Models on Hugging Face](https://huggingface.co/inclusionAI) ## Related Reading Explore more about AI models and technologies: - **[Ling AI Model Family](https://howaiworks.ai/models/ling)** - Complete technical specifications and deployment guide - **[DeepSeek V3.1](https://howaiworks.ai/models/deepseek)** - China's leading open-source AI model - **[Mixture-of-Experts (MoE)](https://howaiworks.ai/glossary/mixture-of-experts)** - Understanding the architecture behind Ling-1T - **[Transformer](https://howaiworks.ai/glossary/transformer)** - The foundational AI architecture - **[Claude Sonnet 4.5](https://howaiworks.ai/models/claude-sonnet)** - Comparison with proprietary models --- *Want to learn more about AI models and their capabilities? Explore our [AI models catalog](https://howaiworks.ai/models) or browse our [glossary of AI terms](https://howaiworks.ai/glossary) for deeper understanding.* ## Frequently Asked Questions ### What is Ling-1T and what makes it special? Ling-1T is Ant Group's flagship AI model with one trillion parameters. It demonstrates state-of-the-art performance in code generation, software development, competitive-level math, and logical reasoning, achieving 70.42% accuracy on AIME 2025. ### What are the three series in the Ling AI model family? The Ling family includes three series: Ling (efficient language models without explicit thinking mode), Ring (thinking models with advanced reasoning), and Ming (multimodal models). ### Is Ling-1T open source? Yes, Ant Group has open-sourced Ling-1T along with Ring-1T-preview, making these trillion-parameter models freely available to the AI community. ### How does Ling-1T perform on mathematical reasoning? On the American Invitational Mathematics Examination (AIME) 2025, Ling-1T achieved 70.42% accuracy while using an average of over 4,000 output tokens per problem, matching industry-leading models. --- Source: https://howaiworks.ai/blog/ant-group-ling-1t-announcement — HowAIWorks.ai