Overview
- Baidu introduced the M100 for inference, targeted for early 2026, and the M300 for both training and inference, slated for early 2027, at its Baidu World event.
- The M100 comes from Baidu’s Kunlunxin unit and is optimized for efficient mixture‑of‑experts inference, while the M300 is designed for ultra‑large multimodal model workloads.
- Baidu’s Tianchi256 supernode, using 256 P800 chips, is planned for the first half of 2026, with a Tianchi512 upgrade using 512 chips set for the second half of 2026.
- The company also unveiled a new Ernie model that handles text, images and video, linking the hardware roadmap to product‑level AI capabilities.
- Executives framed the roadmap as reducing reliance on imported processors under U.S. export restrictions, and Baidu shares rose about 0.8% in Hong Kong morning trading.