Ring

Ring is an open-source trillion-parameter thinking model from Ant Group. It uses the same MoE (Mixture of Experts) architecture as the Ling series, activating about 63B parameters per inference. It is designed for real-world Agent workflows. The model is optimized for coding agents, tool use, and long-horizon task execution, and has achieved leading results on benchmarks such as PinchBench, ClawEval, TAU2-Bench, and GAIA2-search.

Why choose Ring?

Ring is designed for real-world complex tasks. It does not just pursue “smarter” results, but focuses on consistently completing tasks at reasonable cost. In production environments such as coding agents, multi-tool collaboration, engineering development, and research analysis, Ring provides a better balance across quality, speed, and cost.

Built For Real Workflows

Ring-2.6-1T introduces an adjustable Reasoning Effort mechanism, supporting high and xhigh reasoning intensity levels, with adaptive reasoning budget allocation based on task complexity. In tool-intensive, multi-turn Agent pipelines, it can more easily achieve stronger performance with lower token overhead.

• high: Designed for high-frequency Agent workflows, with lower token cost and faster multi-step execution. Suitable for multi-turn interaction, tool collaboration, task decomposition, and production-default invocation.
• xhigh: Designed for high-difficulty tasks such as mathematics, research, complex logical analysis, and multi-path exploration, providing more room for deep reasoning.

This layered reasoning mechanism makes it possible to avoid overthinking simple tasks while fully reasoning on complex tasks, better matching real production task diversity.

Inference Efficiency

The Ring series inherits the sparse activation characteristics of the Ling MoE architecture, activating only a small number of parameters per inference, significantly reducing computational overhead.

• Ring-2.5-1T introduces a linear attention mechanism, reducing memory access overhead by more than 10x and improving generation throughput by over 3x, suitable for deep reasoning and long-cycle task execution.
• Ring-1T is the world’s first open-source trillion-parameter reasoning large model, and the largest and most capable flagship model in the first-generation Ring series.
• Through these optimizations, inference costs are significantly reduced, enabling production-grade deployability.

Long Context Support

Ring-2.6-1T supports a 256K context window, meeting scenarios like long document analysis and multi-step complex reasoning.

• A single request can process long-text input.
• Logic stays consistent across very long reasoning chains, without noticeable performance degradation.
• Supports holistic understanding and generation of large-scale codebases.

Typical Use Cases: Mathematical competition problem solving, code review and generation, academic paper analysis, complex logical reasoning.

Model Details

Ring-2.6-1T

Ring-2.6-1T is a trillion-parameter thinking model with about 63B activated parameters per inference, designed for Agent workflows. It focuses on Agent tasks, tool use, and long-horizon execution, and achieves leading performance on benchmarks such as PinchBench, ClawEval, TAU2-Bench, and GAIA2-search. The model also balances execution quality, latency, and cost, making it suitable for advanced coding agents, complex reasoning pipelines, and large-scale autonomous systems.

Technical Features:

• Scale and efficiency: Trillion-parameter MoE with 63B activated parameters, balancing strong capability and reasoning overhead.
• Adaptive reasoning budget: Two Reasoning Effort levels (high / xhigh) dynamically allocate reasoning resources by task complexity; in tool-intensive, multi-turn scenarios, stronger results can be achieved with lower token overhead.
• high mode (production default): Designed for high-frequency Agent and long-horizon execution, with lower token overhead and faster multi-step execution, suitable for multi-turn interaction, tool collaboration, and task decomposition.
• xhigh mode (high-difficulty tasks): Designed for mathematics, research, complex logical analysis, and multi-path exploration, providing deeper reasoning space.
• Real tasks and benchmarks: Strong performance on execution- and tool-collaboration-oriented evaluations such as PinchBench, ClawEval, TAU2-Bench, and GAIA2-search; demonstrates a higher capability ceiling on difficult reasoning benchmarks such as ARC-AGI-V2, AIME 26, and GPQA Diamond.
• Ecosystem access: Available on OpenRouter  with limited-time free access; will be officially open-sourced later. Note: Ring-2.6-1T is not yet available in the API Console and is currently offered for early access through OpenRouter .

Typical Use Cases:

• Advanced coding agents, programming assistants, and repository-level coding tasks
• Tool orchestration, multi-turn collaboration, and long-horizon autonomous execution
• Complex reasoning pipelines and enterprise-grade workflows
• Scenarios in large-scale autonomous systems that require balancing quality, latency, and cost
• High-difficulty mathematics and research analysis, multi-path reasoning, and complex decision-making

Ring-2.5-1T

Ring-2.5-1T is the previous-generation flagship of the Ring series. Compared to the first-generation Ring-1T, it achieves improvements across generation efficiency, reasoning depth, and long-horizon task execution.

Technical Features:

• Generation Efficiency: Through a high proportion of linear attention mechanism, reduces memory access overhead by more than 10x and improves generation throughput by over 3x when processing sequences exceeding 32K tokens, making it suitable for deep reasoning and long-cycle task execution.
• Deep Reasoning: Introduces a dense reward mechanism based on RLVR, providing feedback constraints on the rigor of the reasoning process, enabling the model to achieve gold medal level in both IMO 2025 and CMO 2025 (self-tested).
• Long-Cycle Task Execution: Through large-scale fully asynchronous agentic reinforcement learning training, significantly enhances long-term autonomous execution capability for complex tasks, with support for agentic programming frameworks like Claude Code and OpenClaw.
• API is now open and can be integrated into production environments.

Typical Use Cases:

• High-difficulty math competition problem solving
• Complex code generation and review
• Academic paper deep analysis
• Multi-step logical reasoning tasks

Evolution History

Version Evolution:

1. Ring 1.0 — Validated the feasibility of training trillion-parameter scale reasoning models with MoE architecture.
2. Ring 2.0 — Introduced FP8 full-precision training, significantly improving training speed and reducing memory overhead.
3. Ring 2.5 — Introduced Hybrid Linear Attention, breaking through the long-context reasoning bottleneck, with end-to-end agentic task execution capability.
4. Ring 2.6 — Introduced adjustable reasoning intensity and adaptive reasoning budget, optimizing coding, tool use, and long-horizon task execution for real-world Agent workflows.

Technology Ecosystem

Based on the Ring reasoning foundation, a complete upstream and downstream technology ecosystem has been further built:

• High-Performance Operator Library: Open-sourced high-performance training and inference operator system, covering the full chain from training to deployment.
• Vertical Domain Models: Industry-specific reasoning models for healthcare, finance, and other sectors, with superior performance in professional task scenarios.
• Open Source Community: All research results are open-sourced to the Inclusion AI  community.

Quick Start

1. Get API Key: Obtain access credentials.
2. Make Your First Call: Complete your first API request in 5 minutes.
3. Explore More Capabilities: Learn more about advanced features like reasoning optimization and long context.