Ring-1T
Ring-1T is Ant Group’s trillion-parameter thinking model, trained from Ling-1T-base through verifiable-reward reinforcement learning and human-feedback alignment. The IcePop stabilization method supports large-scale reinforcement learning, while extended natural-language reasoning targets mathematics, coding, and scientific tasks. Public MIT weights are available with a native 64K context extendable to 128K through YaRN. The original first-party API was retired in June 2026. Earlier preview work is distinct from the October 14 balanced final release recorded in the official announcement.
2025-10-14
1T total, approximately 50B active
Mixture of Experts Transformer with grouped-query attention
MIT
Specifications
- Parameters
- 1T total, approximately 50B active
- Architecture
- Mixture of Experts Transformer with grouped-query attention
- License
- MIT
- Context Window
- 65,536 tokens
- Type
- text
- Modalities
- text
Benchmark Scores
Advanced Specifications
- Model Family
- Ring
- Finetuned From
- Ling-1T-base
- API Access
- Not Available
- Chat Interface
- Not Available
Capabilities & Limitations
- Capabilities
- reasoningcodingtool useinstruction following
- Known Limitations
- Inactive experts still require substantial weight storage.Tool use requires an external execution environment and appropriate chat template.
- Notable Use Cases
- coding agentsenterprise assistantslong-document analysis