Ling-Lite and Ling-Plus
Ant's 290B-total (28.8B active) Ling-Plus MoE trained on lower-spec accelerators, cutting pretraining compute cost by about 20% in its own accounting.
The paper 'Every FLOP Counts' details training a 300B-class MoE without premium GPUs, plus Ling-Lite (16.8B total, 2.75B active). Both were released on Hugging Face.
- Date
- Friday, 7 March 2025
- Lab
- Ant Group (inclusionAI)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total / active parameters (Plus) | 290B / 28.8B | company |
| Compute-cost reduction on lower-spec hardware | about 20% authors' estimate | company |
Date is arXiv v1; Wikipedia says Ling-Plus and Ling-Lite were released in March 2025. The paper does not name the hardware vendors in the abstract I read.
Sources
This record was checked against its sources on 6 October 2026. How we check