Kunlunxin

Baidu unveiled its Tianchi 1024-card super node at the 2026 Baidu World conference, claiming a single unit can train 5-trillion-parameter models and scale horizontally to 10,000-card clusters, with launch planned for next year. The system uses optical engine interconnect for 3x card-to-card bandwidth, delivering 2.5x training and 3x inference performance gains, while a companion G3.5 KV Cache appliance doubles long-context inference throughput and cuts multi-turn TTFT by 90%. The timeline is roughly a year earlier than the 2028 launch announced last year.

Baidu ERNIE disclosed its recursive self-improvement (RSI) system ERNIE Evolve for the first time, which uses a multi-agent dual-loop design to autonomously propose research directions, modify code and validate results, achieving publicly verified SOTA on three metrics across NanoChat AutoResearch and NanoGPT Speedrun. The system has been deployed in real inference optimization for ERNIE 5.5 Flash, lifting generation speed to 623.6% of baseline in low-latency scenarios on Kunlunxin P800 and raising deterministic inference throughput to 133%. Baidu open-sourced ERNIE 5.5 Flash Base on November 5, with a new model expected soon.