数据集 / inclusionAI/Ring-2.5-1T

inclusionAI/Ring-2.5-1T 已完整同步

🤗 Hugging Face   |   🤖 ModelScope    |   🐙 Experience Link Coming Soon~

Ring-2.5-1T，Think Deeper, Run Further

Introducing Ring-2.5-1T: the world's first open-source trillion-parameter thinking model based on hybrid linear attention architecture.

In a major step toward general-purpose AI agents, we're scaling hybrid linear attention across pre-training and RL. Our efficient 1:7 MLA + Lightning Linear Attention boosts reasoning speed and exploration, while expanded RL training enhances deep thinking and long-horizon task execution.

Compared to the previously released Ring-1T, Ring-2.5-1T demonstrates substantial improvements across three key dimensions: generation efficiency, reasoning depth, and long-horizon task execution capabilities：

Generation efficiency: Leveraging a high-ratio linear attention mechanism, Ring-2.5-1T reduces memory access overhead by over 10× and increases generation throughput by more than 3× for sequences exceeding 32K tokens, making it particularly suitable for deep thinking and long-horizon task execution .

Deep Thinking: Building upon RLVR by introducing dense rewards to provide feedback on the rigor of the reasoning process, enabling Ring-2.5-1T to simultaneously achieve gold medal level for both IMO 2025 and CMO 2025 (self-tested).

Long-horizon task Execution: Through large-scale fully-async agentic RL training, significantly enhancing the long-term autonomous execution capability for complex tasks, enabling Ring-2.5-1T to easily adapt to agentic programming frameworks such as Claude Code and the OpenClaw personal AI assistant.

Model Downloads

You can download Ring-1T from the following table. If you are located in mainland China, we also provide the model on ModelScope to speed up the download process.

Model Context Length Download
Ring-2.5-1T 128K -> 256K (YaRN) 🤗 HuggingFace    🤖 ModelScope

Note: If you are interested in the previous version, please visit the past model collections on Huggingface or ModelScope.

Deep Thinking & Long-horizon task Execution

For evaluating the Deep Thinking and Long-term Execution capabilities of Ring-2.5-1T, we selected representative open-source thinking models (DeepSeek-v3.2-Thinking, Kimi-K2.5-Thinking) and closed-source APIs (GPT-5.2-thinking-high, Gemini-3.0-Pro-preview-thinking-high, Claude-Opus-4.5-Extended-Thinking) as references. Ring-2.5-1T achieves state-of-the-art open-source performance across both high-difficulty reasoning tasks—including mathematics, coding, and logical reasoning (IMOAnswerBench, AIME 26, HMMT 25, LiveCodeBench, ARC-AGI-V2)—and long-horizon task execution such as agent search, tool calling, and software engineering (Gaia2-search, Tau2-bench, and SWE-Bench Verified).

We also conducted additional tests on the "heavy thinking mode," by expanding parallel thinking and summarization during the reasoning process to achieve test-time scaling, thereby effectively enhanc

171 个文件

浏览文件

数据集版权信息

本数据集的许可证为 MIT License。如有违反相关条款，请联系 WEHUB，我们将及时处理。 查看许可证