Ling-1T
First trillion-parameter flagship of Ant's non-thinking Ling 2.0 line, with about 50B active parameters, 20T+ tokens, 128K context and an MIT licence.
Uses evolutionary chain-of-thought across mid- and post-training and a syntax-function-aesthetics reward for front-end code. Ant says it tops open models on ArtifactsBench. Companion paper arXiv 2510.22115.
- Date
- Thursday, 2 October 2025
- Lab
- Ant Group (inclusionAI)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters | 1T / ~50B | company |
| Pretraining tokens | 20T+ | company |
Date is HF weights upload (2025-10-02 UTC); I found no dated announcement post. Tech report v1 2025-10-25. Benchmark claims are Ant's.
Sources
This record was checked against its sources on 6 October 2026. How we check