Ring-1T

Ring-1T is an open-weight reasoning model released in October 2025 by inclusionAI, the open-source research initiative associated with Ant Group, the Alibaba-affiliated financial-technology company that operates the Alipay payments platform. With roughly 1 trillion total parameters and only about 50 billion activated per token through a sparse mixture-of-experts (MoE) design, Ring-1T was described by its developers and by independent observers as the first open-source, trillion-parameter-scale "thinking" model, making it one of the largest openly released reasoning models at the time.^[1][2] It is the reasoning ("thinking") flagship of inclusionAI's paired Ling and Ring model families, both built on a shared "Ling 2.0" architecture.^[3]

Overview

Ring-1T is a chain-of-thought reasoning model: it produces an extended internal deliberation before emitting a final answer, a design pattern popularized by models such as OpenAI's o1 and DeepSeek-R1. It is built on the same trillion-parameter foundation as inclusionAI's non-thinking model, Ling-1T, and was further trained with large-scale reinforcement learning to specialize in mathematics, code generation, and logical reasoning.^[1][3]

The model was first released as "Ring-1T-preview" on September 30, 2025, with the complete Ring-1T following in mid-October 2025; the accompanying technical report, titled "Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model," was published on arXiv on October 21, 2025.^[1][2] Weights were distributed openly on Hugging Face and ModelScope under the permissive MIT license, including a quantized FP8 variant to ease deployment of the very large model.^[2]

Specifications

Developer: inclusionAI and Ant Group

inclusionAI is the open-source brand under which Ant Group releases its large language models. Ant Group, formerly Ant Financial, is an Alibaba-affiliated company best known for operating Alipay, one of China's largest digital payment and financial-services platforms. Ant began open-sourcing its Ling MoE models under the inclusionAI umbrella in 2025, positioning itself as a new entrant among China's frontier-model developers alongside DeepSeek, Alibaba's Qwen team, and Moonshot AI, maker of Kimi.^[3][4] The teams behind the models are sometimes referred to internally by the "Bailing" (Ling) name, and the technical report credits a large "Ling Team" of more than a hundred contributors.^[1]

The Ling and Ring families

inclusionAI organizes its models into two parallel lines that share a common foundation. The Ling series consists of general-purpose, non-thinking instruction models, while the Ring series consists of reasoning ("thinking") models that perform explicit step-by-step deliberation. Both are built on the Ling 2.0 mixture-of-experts architecture, which inclusionAI began open-sourcing in September 2025 starting with smaller variants such as Ling-mini-2.0 and Ring-mini-2.0.^[3][5]

At the top of the range, the non-thinking Ling-1T and the thinking Ring-1T share the same trillion-parameter MoE backbone: Ling-1T is the flagship instruction model, and Ring-1T is the reasoning model trained on top of the Ling-1T-base foundation.^[3] Ling-1T introduced architectural choices intended to stabilize training at trillion scale, including an auxiliary-loss-free, sigmoid-scored expert-routing balance strategy and QK normalization, which Ring-1T inherits.^[6] inclusionAI later iterated on the line with updated releases such as the Ling/Ring 2.5 and 2.6 generations in 2026, the latter of which increased activated parameters and extended the context window.^[7]

Architecture and RL training

Ring-1T's sparse MoE design means that although the model contains on the order of 1 trillion parameters in total, only a small fraction, about 50 billion, is activated for any given token. This lets the model approach the capability of a very large dense network while keeping per-token inference cost far lower than its nominal size would suggest. The model supports a context window of up to 128K tokens.^[1][2]

The technical report frames Ring-1T primarily as an exercise in scaling reinforcement learning to trillion-parameter models, and it introduces several methods to make that feasible:^[1]

• IcePop (sometimes styled "Icepop") stabilizes RL training by addressing the discrepancy between the training and inference computation paths in MoE models. The report describes it as using token-level discrepancy masking and clipping (a "masked bidirectional truncation" technique) to suppress instability that would otherwise accumulate over long training runs.^[1][2]
• C3PO++ improves efficiency for long reasoning rollouts by dynamically partitioning them, improving time and resource utilization during generation-heavy RL.^[1]
• ASystem is a high-performance RL framework built to overcome the systemic engineering bottlenecks of trillion-parameter training. It uses a SingleController plus SPMD design with features such as transparent memory offloading and a high-throughput, sandboxed reward service.^[1][2]

Training combined reinforcement learning with verifiable rewards (RLVR), where correctness can be checked automatically for tasks like math and code, with reinforcement learning from human feedback (RLHF) for broader alignment.^[2]

Benchmarks

inclusionAI reported that Ring-1T achieves leading open-source results across math, code, and logical-reasoning benchmarks, and approaches the performance of leading closed (proprietary) systems on several tasks. Reported scores include 93.4 on AIME 2025 and 86.72 on HMMT 2025 (both competition-mathematics benchmarks), a CodeForces rating of 2088, and 55.94 on the abstract-reasoning benchmark ARC-AGI-1.^[2][8] The developers also cited strong results on LiveCodeBench, the comprehensive Arena-Hard v2.0 benchmark, the HealthBench medical benchmark, and Creative Writing v3, and positioned Ring-1T against open models such as DeepSeek-V3.1 and Qwen3-235B-A22B-Thinking as well as closed APIs including Gemini 2.5 Pro and GPT-5 (Thinking).^[1][8]

The most widely cited result is Ring-1T's performance on the 2025 International Mathematical Olympiad (IMO). According to inclusionAI, the model produced correct solutions to problems 1, 3, 4, and 5, reaching a silver-medal-level standard, with a near-complete solution to problem 2; it did not solve problem 6.^[1][2] On the 2025 International Collegiate Programming Contest (ICPC) problem set, the model was reported to solve 5 of 11 problems within three attempts.^[2] As with all self-reported benchmark figures, these results come from the developer's own evaluation and had not been independently standardized at release; they should be read as the developer's claims rather than neutral measurements.

Availability and significance

Ring-1T was released as an open-weight model under the MIT license on Hugging Face and ModelScope, with an FP8 quantized version provided to reduce the memory needed to run it.^[2] Even so, the model's trillion-parameter scale places it beyond the reach of typical consumer hardware, and it is generally accessed through cloud inference providers or substantial multi-GPU deployments; third-party platforms such as SiliconFlow began hosting it shortly after release.^[8]

Ring-1T is significant as one of the first openly released reasoning models at trillion-parameter scale, and as a marker of Ant Group's entry into frontier model development. Its arrival extended a broader 2025 trend of Chinese laboratories releasing very large open-weight MoE systems, alongside DeepSeek's V3 and R1 models, Alibaba's Qwen series, and Moonshot AI's Kimi K2, narrowing the gap between open models and leading proprietary systems on reasoning-heavy tasks.^[3][4] The work also drew attention for its emphasis on the reinforcement-learning infrastructure and stabilization techniques (IcePop, C3PO++, and ASystem) needed to apply large-scale RL to models of this size, contributions that the team published in detail rather than keeping proprietary.^[1]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12inclusionAI / Ling Team. "Every Step Evolves: Scaling Reinforcement Learning for Trillion-Scale Thinking Model." arXiv:2510.18855, October 21, 2025. arxiv.org/...2510.18855
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11inclusionAI. "Ring-1T" model card. Hugging Face. huggingface.co/...Ring-1T
3. ^1 ^2 ^3 ^4 ^5 ^6inclusionAI / AI Wiki. "InclusionAI." aiwiki.ai/...inclusionai
4. ^1 ^2Interconnects. "Inside a Chinese frontier lab: interview on Ant Group's open model ambitions." interconnects.ai/...chinese-frontier-lab-inclusion
5. ^Ant Ling. "Ring-mini-2.0: Small Model, Great Intelligence." Medium. ant-ling.medium.com/...t-intelligence-38040fdf5dff
6. ^Ant Ling. "Deep Insight, Efficient Inference: Introducing the Trillion-Parameter Ling-1T Model." Medium. ant-ling.medium.com/...-ling-1t-model-77d6170e5e8e
7. ^Codersera. "Ring-2.6-1T: Ant Group's Open Trillion-Parameter Reasoning Model." 2026. codersera.com/...on-parameter-reasoning-model-2026
8. ^1 ^2 ^3SiliconFlow. "Ring-1T Now on SiliconFlow: The World's First Open-Source Trillion-Parameter Thinking Model." siliconflow.com/...illion-parameter-thinking-model