Baidu Tianchi 1024-card super node unveiled: can support training of 5-trillion-parameter models, coming to market next year

IT Home onsite news, November 12: At the opening ceremony of today's Baidu World 2026 conference, Baidu Group Executive Vice President Shen Dou unveiled the Baidu Tianchi 1024-card super node. This computing cluster will be brought to market next year.

The Baidu Tianchi 1024-card super node features industry-leading optical engine technology, with inter-card interconnect bandwidth increased 3x, enabling high-speed data interconnection. Compared with the previous generation product, training performance is improved 2.5x and inference performance 3x.

The Baidu Tianchi 1024-card super node is paired with the G3.5 KV Cache memory acceleration cabinet, strengthening cache reuse capabilities. Long-context inference performance is improved 2x; in multi-turn dialogue cache-hit scenarios, TTFT (time to first token) is reduced by 90%, and overall throughput is increased 7x.


A single Tianchi 1024-card super node can support training of large models with 5 trillion parameters and offers excellent horizontal scaling capability, able to scale to clusters of ten thousand cards.
IT Home notes that at last year's Baidu World conference, Baidu announced that the Tianchi thousand-card-class super node would come to market in 2028; the timeline has now been moved up by a year. In addition, the Kunlunxin N series has also been moved up to a 2028 release.
