ERNIE Lightning
Baidu released ERNIE Lightning, a reasoning model built for Baidu Search, on its Qianfan platform with API access for enterprise customers and partners.
Baidu reworked the standard Transformer with BSA (Baidu Sparse Attention), a sparse attention method based on feedback-driven KV-cache compression. A fast-thinking model serves live requests while a slow-thinking model acts as a judge that feeds back improvements. It has thinking and non-thinking modes. Baidu says it runs fast inference on a single eight-card Kunlunxin server and cuts inference engine cost by more than 30%. Baidu also released the video model ERNIE-Wizshot-ME the same day, which turns still portraits into 4 to 15 second 1080p clips at 0.7 yuan per second.
- Date
- Thursday 12 November 2026
- Lab
- Baidu
- Kind
- model
- Access
- closed API
Figures
| Measure | Value | Measured by |
|---|---|---|
| Inference engine cost reduction | more than 30% with Kunlunxin co-optimization | company |
| ERNIE-Wizshot-ME list price | 0.7 yuan per second 4 to 15 second 1080p videos | company |
Announced at Baidu World 2026. Baidu has published no independent benchmarks, parameter counts or context lengths. ERNIE-Light-DT-1.0 on Qianfan and OpenRouter may or may not be the same product.
Sources
This record was checked against its sources on 12 November 2026. How we check