BitNet b1.58 2B4T
First open native 1-bit LLM at scale, with 2B parameters trained on 4T tokens and matching similar-size full-precision open models, per the authors.
Moved BitNet from small experiments to a usable open model with GPU and CPU inference code, claiming lower memory, energy and decoding latency than comparable full-precision models.
- Date
- Wednesday, 16 April 2025
- Lab
- Microsoft
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Scale | 2B params, 4T tokens weights on Hugging Face; report labelled work in progress | company |
Sources
This record was checked against its sources on 6 October 2026. How we check