Source量子位•Stalecollected in 46m
Ant Group Bailing Ring-2.6-1T Enhances Agent Capabilities
New open-source agent model hits 95.83 on AIME 26, offering high-tier reasoning performance.
30-Second TL;DR
What Changed
Bailing Ring-2.6-1T model released with improved agent execution
Why It Matters
This release provides developers with a high-performing open-source option for building agentic systems, challenging existing benchmarks.
What To Do Next
Download and benchmark the Bailing Ring-2.6-1T model on your specific agentic tasks to compare against current GPT-4 or Claude models.
Who should care:Researchers & Academics
Key Points
- •Bailing Ring-2.6-1T model released with improved agent execution
- •Achieved 95.83 score on AIME 26 benchmark
- •Focus on open-source accessibility for agentic workflows
Deep Insight
Background and context from public sources — not the original article. 10 sources cited.
Enhanced Key Takeaways
- •Ant Group's Bailing Ring-2.6-1T is a trillion-parameter model with 63 billion activated parameters, designed for complex tasks and production environments, and features a novel 'Dynamic Thinking Intensity' mechanism.
- •The model operates in two distinct modes: an 'Agent mode (high)' optimized for multi-step execution and tool invocation, and a 'deep reasoning mode (xhigh)' tailored for mathematical reasoning and scientific research.
- •In its 'Agent mode (high)', Ring-2.6-1T achieved a PinchBench score of 87.60, outperforming GPT-5.4 xHigh and Gemini-3.1-Pro high, and also scored 63.82 on ClawEval.
- •Ring-2.6-1T is part of Ant Group's broader open-source 'BaiLing' (also known as Ling) model family, which includes Ling (general language models), Ring (reasoning models), and Ming (multimodal systems).
- •Ant Group's open-source AI strategy, exemplified by the Bailing family, aims to accelerate the integration of AI into real-world applications and build trusted, open-source infrastructure for the AI era, extending beyond its traditional financial services applications.
Competitor Analysis
Ant Group Bailing Ring-2.6-1T
- AIME 26 Score
- 95.83
- PinchBench Score (Agent Mode)
- 87.60 (surpasses GPT-5.4 xHigh, Gemini-3.1-Pro high)
- GPQA Diamond Score (Deep Reasoning Mode)
- 88.27
- License
- Open Source
Moonshot AI Kimi K2.6
- AIME 26 Score
- 96.4%
- PinchBench Score (Agent Mode)
- N/A
- GPQA Diamond Score (Deep Reasoning Mode)
- N/A
- License
- Proprietary
Alibaba Cloud / Qwen Team Qwen3.6 Plus
- AIME 26 Score
- 95.3%
- PinchBench Score (Agent Mode)
- N/A
- GPQA Diamond Score (Deep Reasoning Mode)
- N/A
- License
- Open Source
Zhipu AI GLM-5.1
- AIME 26 Score
- 95.3%
- PinchBench Score (Agent Mode)
- N/A
- GPQA Diamond Score (Deep Reasoning Mode)
- N/A
- License
- Proprietary
Anthropic Claude Opus 4.6
- AIME 26 Score
- 87.2%
- PinchBench Score (Agent Mode)
- N/A
- GPQA Diamond Score (Deep Reasoning Mode)
- 89.4%
- License
- Proprietary
| Model | AIME 26 Score | PinchBench Score (Agent Mode) | GPQA Diamond Score (Deep Reasoning Mode) | License |
|---|---|---|---|---|
| Ant Group Bailing Ring-2.6-1T | 95.83 | 87.60 (surpasses GPT-5.4 xHigh, Gemini-3.1-Pro high) | 88.27 | Open Source |
| Moonshot AI Kimi K2.6 | 96.4% | N/A | N/A | Proprietary |
| Alibaba Cloud / Qwen Team Qwen3.6 Plus | 95.3% | N/A | N/A | Open Source |
| Zhipu AI GLM-5.1 | 95.3% | N/A | N/A | Proprietary |
| Anthropic Claude Opus 4.6 | 87.2% | N/A | 89.4% | Proprietary |
Technical Deep Dive
- Ring-2.6-1T is a trillion-parameter model with 63 billion activated parameters.
- It incorporates a 'Dynamic Thinking Intensity' mechanism, allowing it to flexibly balance cognitive depth, token cost, and execution speed based on computational demands.
- The model offers two operational modes: 'high' (Agent mode) for multi-step execution and tool invocation, and 'xhigh' (deep reasoning mode) for mathematical reasoning and scientific research.
- It is optimized for coding agents, tool use, and long-horizon task execution.
- The model features a context window of 262,144 tokens.
- Its underlying architecture is a Mixture-of-Experts (MoE) with hybrid attention, designed to handle long contexts efficiently.
- Ant Group has developed an asynchronous RL training system with an 'ice pop algorithm' to enhance the stability of trillion-scale reinforcement learning for this model.
Future ImplicationsAI analysis grounded in cited sources
Ant Group's open-sourcing of advanced agent models will significantly accelerate the adoption of AI agents in financial and other industries.
By providing highly capable, open-source models like Ring-2.6-1T, Ant Group lowers the barrier for developers and enterprises to build and deploy sophisticated AI agents, particularly in fintech where Ant Group has deep roots and an established ecosystem.
The 'Dynamic Thinking Intensity' mechanism introduced in Ring-2.6-1T will influence future designs of efficient agentic AI models.
This mechanism directly addresses the critical trade-off between computational cost and reasoning depth, offering a practical solution for deploying powerful agents in diverse, real-world scenarios with varying resource constraints, thus setting a potential new standard for efficiency.
Ant Group will solidify its position as a major contributor to the global open-source AI landscape, particularly in agentic and reasoning models.
The consistent release of high-performing, trillion-parameter open-source models under the Ling/Ring/Ming family, coupled with Ant Group's stated commitment to open-source and Artificial General Intelligence (AGI), indicates a long-term strategic investment and growing influence in the field.
Timeline
2023-11
Ant Group's Bailing large-scale model completed regulatory record-filing in China, enabling public release of products.
2025-07
Ant Digital Technology launched Agentar-Fin-R1, a large financial reasoning model.
2025-09
Ant Group open-sourced Ring-1T-preview, a trillion-parameter reasoning model.
2025-10
Ant Group unveiled the Ling 2.0 series, encompassing Ling, Ring, and Ming models.
2026-02
Ant Group released Ling-2.5-1T and Ring-2.5-1T, alongside the multimodal Ming-Flash-Omni-2.0, as a comprehensive upgrade to its open-source family.
2026-05
Ant Group launched Ring-2.6-1T, a trillion-parameter flagship reasoning model with enhanced agent capabilities.
- 2023-11Ant Group's Bailing large-scale model completed regulatory record-filing in China, enabling public release of products.
- 2025-07Ant Digital Technology launched Agentar-Fin-R1, a large financial reasoning model.
- 2025-09Ant Group open-sourced Ring-1T-preview, a trillion-parameter reasoning model.
- 2025-10Ant Group unveiled the Ling 2.0 series, encompassing Ling, Ring, and Ming models.
- 2026-02Ant Group released Ling-2.5-1T and Ring-2.5-1T, alongside the multimodal Ming-Flash-Omni-2.0, as a comprehensive upgrade to its open-source family.
- 2026-05Ant Group launched Ring-2.6-1T, a trillion-parameter flagship reasoning model with enhanced agent capabilities.
Sources (10)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
1Google Search Sourcevertexaisearch.cloud.google.com2Google Search Sourcevertexaisearch.cloud.google.com3Google Search Sourcevertexaisearch.cloud.google.com4Google Search Sourcevertexaisearch.cloud.google.com5Google Search Sourcevertexaisearch.cloud.google.com6Google Search Sourcevertexaisearch.cloud.google.com7Google Search Sourcevertexaisearch.cloud.google.com8Google Search Sourcevertexaisearch.cloud.google.com9Google Search Sourcevertexaisearch.cloud.google.com10Google Search Sourcevertexaisearch.cloud.google.com
Weekly AI Recap
Read this week's curated digest of top AI events →
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 量子位 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
The weekly digest
One email a week. Unsubscribe anytime.