Ring-2.6-1T Brings Trillion-Parameter Reasoning to Agentic Workflows

By vramkickedin | | 2 min read
A digital ring hovers to the right of view formed from billions of tiny interconnected nodes.

Ring-2.6-1T is a newly released trillion-parameter reasoning model designed for complex, multi-step tasks in real-world settings. It moves beyond simple question answering to handle continuous agent workflows, tool use, and long-horizon planning. Developers and researchers can now download the model for validation, adaptation, and further development.

InclusionAI, the team behind Ling-2.6-1T and Ming-Flash-Omni-2.0 built Ring-2.6-1T to serve production environments like enterprise automation, engineering, and scientific analysis. Instead of just scaling parameters, the team focused on making the model execute tasks stably and adapt its reasoning effort. The result is a flexible tool for anyone who needs AI to drive real workflows rather than generate isolated answers.

Adjustable reasoning and agent execution

Key Features
  • Stable multi-step task and tool collaboration
  • Two reasoning effort levels: high and xhigh
  • Asynchronous RL training with IcePop algorithm
  • Context length up to 256K via YaRN
  • Outperforms GPT-5.4 on PinchBench agent benchmark
  • Scores 95.83 on AIME 26 math benchmark

Enterprise teams and developers building coding agents, process automation, or scientific research tools will find this release most useful. The high and xhigh reasoning modes let users trade speed for depth depending on task complexity. Privacy-conscious shops with the necessary GPU infrastructure can run the model locally, giving them full control over their data and costs.

Deployment and what’s next

Running Ring-2.6-1T requires significant compute, with the reference deployment using SGLang across multiple GPU nodes. The model weights are available in BF16 and FP8 formats, and a separate download is provided for users in mainland China via ModelScope. inclusionAI also plans to release a technical report detailing the Stick-Breaking algorithm used in the asynchronous RL training pipeline.

“Ring-2.6-1T not only understands user intent but can also continuously drive tasks forward in real workflows.” — Source: Hugging Face

Project Links