Released by InclusionAI (an AGI initiative by Ant Group), Ring-2.6-1T is a trillion-parameter reasoning model specifically optimized for complex agentic workflows and long-horizon task execution. It serves as the reasoning-focused counterpart to the Ling-2.6 series, utilizing a sparse Mixture-of-Experts (MoE) architecture that activates approximately 63 billion parameters per token. The model was trained using an innovative asynchronous reinforcement learning paradigm combined with the IcePop algorithm, which enhances training stability and performance for trillion-scale reasoning pipelines.
A defining feature of Ring-2.6-1T is its Reasoning Effort mechanism, which allows developers to toggle between two intensity levels: high and xhigh. The "high" mode is tailored for high-frequency agent interactions and multi-step tool collaboration, aiming for efficiency and reduced token overhead. The "xhigh" mode is designed for the most challenging reasoning tasks, such as competitive mathematics and scientific research, where the model performs extensive internal exploration before outputting a solution.
The model architecture incorporates a hybrid of Multi-Head Latent Attention (MLA) and linear attention to maintain high throughput and a manageable VRAM footprint across its large context window. It natively supports explicit chain-of-thought (CoT) traces, emitting internal reasoning within <think>...</think> tags to provide transparency into its decision-making process. This behavior is particularly useful for debugging autonomous agents and verifying the logic of complex code generations.
In performance evaluations, Ring-2.6-1T has demonstrated state-of-the-art results on several benchmarks, including a score of 95.83 on AIME 2026 and 88.27 on GPQA Diamond. It has also shown high reliability in real-world software engineering tasks, ranking at the top of the SWE-bench Verified and ArtifactsBench leaderboards. For developers, the model is compatible with major inference engines like vLLM and SGLang, and it supports custom chat templates that include a reasoning_effort parameter for dynamic control over thinking depth.