Ant Group released two trillion-parameter open-source AI models, Ling-2.5-1T and Ring-2.5-1T, expanding its Ling family and signaling progress in efficient reasoning and multimodal capabilities for financial and digital services.

Both models are released under open licenses and are available on Hugging Face and ModelScope, continuing Ant Group’s strategy of open-source AI distribution to researchers and developers.

This release follows Ling 2.0, unveiled in late 2025, and marks ongoing expansion of Ant Group’s open-source AI ecosystem.

On AIME 2026, Ling-2.5-1T matches frontier models that typically require 15k–23k tokens using about 5,890 tokens, illustrating improved efficiency.

The announcements fit a broader fintech trend of building proprietary AI infrastructure and integrating large-language and reasoning systems into finance tasks like compliance, customer interactions, and risk assessment.

A coordinated strategy across Ling, Ring, and Ming signals a push to support enterprise AI deployments that combine language understanding, reasoning, and multimodal interactions in financial services.

Ring-2.5-1T is a reasoning-optimized model with a hybrid linear architecture, delivering strong mathematical results on IMO 2025 (35/42) and CMO 2025 (105/126).

Ring-2.5-1T, as part of Ant Group’s Ring series, targets advanced mathematical reasoning with strong benchmarks (IMO and CMO) using a hybrid linear design.

The Ring series extends with Ring-2.5-1T, boasting gold-tier performance on major math competitions and optimized for higher-order reasoning.

Open-release approach is intended to accelerate ecosystem adoption by letting external developers build domain-specific finance applications, potentially reducing costs and enabling larger-scale deployments.

Open access positions Ant Group among global contributors in open large-scale AI and supports external developers in adapting models for financial-domain tasks.

Ling-2.5-1T is a flagship general language model designed for efficient reasoning and agent interaction, supporting context lengths up to one million tokens and achieving frontier-like results with about 5,890 tokens.