InclusionAI

inclusionAI
Type	Research division
Industry	Artificial intelligence
Founded	2024
Headquarters	Hangzhou, China
Key people	He Zhengyu (CTO of Ant Group)
Parent	Ant Group
Owner	Ant Group
Products	Large Language Models Reinforcement Learning systems AGI frameworks Multimodal models
Website	inclusionai.github.io

inclusionAI is an artificial general intelligence (AGI) research initiative established by Ant Group, focused on developing and open-sourcing advanced artificial intelligence systems. The initiative represents Ant Group's dedicated effort to work towards AGI through the development of Large Language Models (LLMs), Reinforcement Learning (RL) systems, multimodal models, and other AI-related frameworks and applications. inclusionAI describes itself as a hub for open projects from Ant Group's research teams working toward reproducible and community-driven AI systems, with a stated mission to develop a fully open-sourced AI ecosystem.

Overview

inclusionAI operates as the primary vehicle for Ant Group's artificial general intelligence ambitions, maintaining a strong commitment to open-source principles and collaborative development. The organization develops and releases various AI models and tools designed to advance the field of AGI while ensuring accessibility and inclusivity in AI development.

The initiative is guided by principles of fairness, transparency, and collaboration, with a focus on tools for training and evaluating reasoning-oriented LLMs via RL, agent frameworks, and the release of trained model checkpoints when feasible. This aligns with Ant Group's broader "AI First" corporate strategy announced in 2024. Public materials indicate that inclusionAI maintains repositories on GitHub and model artifacts on Hugging Face, and has presented work at venues such as ICLR 2025 Expo.

History

inclusionAI emerged as part of Ant Group's increased focus on artificial intelligence research and development. The initiative became prominently active in 2024-2025 with the release of multiple open-source models and frameworks. This move aligned with Ant Group's "Plan A" recruitment initiative, launched in April 2025, which aimed to attract top AI talents and ramp up innovation efforts.

By May 2025, Ant Group publicly showcased its elite AI researchers, including figures like He Zhengyu, a PhD graduate from the Georgia Institute of Technology known for developing advanced algorithms. Public references to inclusionAI as a named project appear in 2025 in connection with an ICLR Expo session highlighting its open RL training stack and agent work.

In March 2025, Ant Group announced the open-sourcing of the Ling Mixture of Experts (MoE) Large Language Models under the inclusionAI umbrella, marking a significant milestone in the initiative's development. This was followed by the release of Ling-Plus and Ling-Lite models, which demonstrated the ability to train large-scale models on domestically produced Chinese chips from Alibaba and Huawei.

inclusionAI's projects began appearing on platforms like GitHub and Hugging Face in mid-to-late 2025, with releases such as the Inclusion Arena leaderboard in August 2025. In September 2025, the organization began open-sourcing Ling 2.0, a series of MoE architecture LLMs, with Ling-mini-2.0 as the first released version. On September 30, 2025, the organization released Ring-1T-preview, a trillion-parameter reasoning model.

Products and Models

Large Language Models

inclusionAI has developed multiple families of LLMs with a focus on efficiency, reasoning capabilities, and multimodal processing:

Model Family	Description	Key Features
Ling Series	Foundation LLMs with MoE architecture	Ling-Lite: 16.8B parameters (2.75B activated) Ling-Plus: 290B parameters (28.8B activated) Ling-mini-2.0: 16B parameters (1.4B activated) Ling-V2: Enhanced version with improved capabilities
Ring Series	Reasoning-focused LLMs	Ring-V2: Reasoning MoE LLM Ring-lite-2507: 16.8B MoE model with 2.75B activated parameters Ring-1T-preview: Trillion-parameter model (preview checkpoint)
Ming Series	Multimodal LLMs	Ming-lite-omni: Multimodal understanding and generation Ming-lite-omni 1.5: Enhanced multimodal capabilities Ming-Omni: Advanced multimodal model

Ring-1T-preview

Ring-1T-preview is a preview checkpoint of a trillion-parameter "thinking" model released in late September 2025 on Hugging Face. The model features a MoE architecture and was positioned to facilitate early community exploration. It excels in natural language reasoning and was trained on 20 trillion tokens, achieving 92.6% on the AIME 2025 (American Invitational Mathematics Examination) math benchmark. The model is optimized for tasks requiring deep thinking and long-term planning, such as code generation and complex problem-solving, and supports long-horizon problem solving.

The model was fine-tuned using inclusionAI's custom RLVR framework with the icepop method. FP8 variants and community quantizations appeared shortly after on Hugging Face. Third-party coverage reported Ring-1T-preview as the first open-source trillion-parameter model.

Ming-Omni

Ming-Omni is an advanced open-source multimodal model capable of processing images, text, audio, and video, released in 2025. The model features a comprehensive multimodal processing architecture with MoE design and modality-specific routers. It supports speech and image generation, dialect understanding, voice cloning, context-aware dialogues, text-to-speech, and image editing.

Ming-Omni represents a breakthrough in multimodal AI, integrating dedicated encoders for different modalities. It supports a wide range of tasks without additional fine-tuning, including generating natural speech, high-quality images, and handling dialect-specific interactions. The model has been described as the first open-source model matching GPT-4o's modality support, with all code and weights publicly available.

Frameworks and Tools

inclusionAI has developed several frameworks to support AGI research and development:

AReaL (Ant Reasoning RL)

AReaL is an open-source, fully asynchronous reinforcement learning training system designed for large reasoning and agentic models. It decouples generation from training to improve GPU utilization and training stability, and provides details intended for full reproducibility (data, infra, and models). The system emphasizes lightning-fast, efficient operations for training large-scale models, and was developed by the AReaL Team at Ant Group in collaboration with Tsinghua University's Institute for Interdisciplinary Information Sciences.

ASearcher

ASearcher is an open-source framework for large-scale online RL training of search agents, aiming to advance "Search Intelligence" to expert-level performance. The framework offers guidance to build customized agents, including integration with AReaL.

AWorld

AWorld is a runtime system for building, evaluating and training general multi-agent assistance. The system provides infrastructure for developing collaborative agent systems and testing their performance in various scenarios.

Inclusion Arena

Inclusion Arena is a live leaderboard and open platform for evaluating large foundation models based on real-world, in-production applications, launched in August 2025. The platform bridges AI-powered apps with state-of-the-art LLMs and multimodal LLMs (MLLMs).

Unlike traditional lab-based benchmarks, Inclusion Arena prioritizes evaluations based on production environments to better reflect practical utility and addresses gaps in conventional evaluation methods by using production data. The platform was proposed by researchers from inclusionAI and Ant Group and shifts the paradigm of model evaluation from synthetic lab benchmarks to real-world performance metrics derived from production applications. The platform is live and open, inviting contributions from the AI community.

ABench

ABench is a benchmark suite for evaluating AI models developed by inclusionAI.

Key Repositories and Releases

Project	Type	First Public Reference/Release	Primary Link
AReaL	RL training system for LLM reasoning	2025 (paper + repo updates)	GitHub
ASearcher	RL system for search agents	2025	GitHub
AWorld	Multi-agent assistance runtime	2025	GitHub
Ring-1T-preview	Trillion-parameter model (preview checkpoint)	September 2025	Hugging Face
Ming-Omni	Advanced multimodal model	2025	Project Page
Inclusion Arena	Live evaluation leaderboard	August 2025	arXiv

Technical Approach and Innovations

Open, Reproducible Systems

inclusionAI's work emphasizes (i) open, reproducible RL training pipelines for reasoning-centric LLMs; (ii) asynchronous system designs that reduce training latency bottlenecks by decoupling rollout generation from parameter updates; and (iii) releasing code, data notes, and, when feasible, model weights for community use and inspection.

Cost-Efficient Training

inclusionAI has pioneered methods for training large-scale models on resource-constrained hardware. The organization reported training costs of approximately $880,000 for their Ling models, representing a 20% cost reduction compared to traditional approaches. This was achieved through:

• Use of domestically produced Chinese chips from Alibaba and Huawei
• Implementation of the EDiT (Elastic Distributed Training) method
• FP8 mixed-precision training throughout the entire process
• Novel optimization techniques for heterogeneous computing environments

Open Source Commitment

All major models and frameworks developed by inclusionAI are released as open-source software, available through platforms including:

• GitHub (primary repository)
• Hugging Face (model distribution)
• ModelScope (Chinese platform)

Research Focus Areas

inclusionAI's research spans multiple domains critical to AGI development:

• Natural language processing and understanding
• Reinforcement learning for reasoning and agent systems
• Multimodal AI combining vision, language, and speech
• Efficient training methods for resource-constrained environments
• Agent-based systems and multi-agent coordination
• Real-world evaluation and benchmarking

Collaboration and Community

The initiative actively encourages collaboration from researchers, developers, and AI enthusiasts worldwide. inclusionAI maintains:

• Open-source repositories with over 2,000 projects
• Active presence on developer platforms
• Integration with Ant Group's broader AI ecosystem
• Partnerships with academic institutions like Tsinghua University
• Collaborations with industry researchers

Relationship to Ant Group and Ecosystem

inclusionAI sits within the wider Ant Group technology and open-source ecosystem, which spans databases, privacy computing, and AI infrastructure. As part of Ant Group, inclusionAI's work supports the parent company's broader AI initiatives, including:

• Healthcare AI applications through the AQ app
• Financial services AI solutions
• Integration with Alipay and other Ant Group services

The models developed by inclusionAI are planned for use in industrial AI solutions across healthcare, finance, and other sectors served by Ant Group. Ant Group communicates its open-source and research activities through corporate channels and events such as the INCLUSION·Conference on the Bund in Shanghai, where it shares AI initiatives and related reports.

In September 2025, at the INCLUSION·Conference on the Bund, Ant Group highlighted its AI advancements, including open-source contributions from inclusionAI, underscoring the initiative's role in promoting trustworthy AI across industries.

See also

• Ant Group
• Artificial general intelligence
• Large language model
• Mixture of experts
• Open-source artificial intelligence
• Reinforcement learning
• Multimodal learning
• GitHub
• Hugging Face
• DeepSeek

External links

• Official website
• GitHub organization
• Hugging Face profile
• Twitter/X account