Ring-2.5-1T
Billed as the first open trillion-parameter thinking model on hybrid linear attention (1:7 MLA to Lightning Linear); gold-level on IMO 2025 and CMO 2025 (self-tested).
The linear-attention mix cuts memory access by over 10x and raises throughput over 3x beyond 32K tokens; dense process rewards and fully async agentic RL target long-horizon work. MIT.
- Date
- Tuesday, 10 February 2026
- Lab
- Ant Group (inclusionAI)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Memory-access reduction (beyond 32K tokens) | over 10x company-claimed | company |
| Generation throughput gain | over 3x company-claimed | company |
Date from HF weights upload; Ling-2.5-1T followed on 2026-02-14 with 63B active parameters, 29T training tokens and 1M context. 'First' is Ant's claim; competing hybrid-linear releases include Kimi Linear (2025-10-30).
Sources
This record was checked against its sources on 6 October 2026. How we check