Baidu reports AI agent accelerated development model’s performance
The company says ERNIE Evolve improved throughput by 33.7% on GPUs, while a larger gain on Kunlunxin chips involved previously unused features.
Baidu has introduced ERNIE Evolve, a system that conducts AI coding experiments, reporting performance gains for its ERNIE-5.5-Flash model, which remains under development. The company said throughput on engineer-optimized GPUs increased by 33.7% over about three days. The performance figures are all supplied by Baidu.
The system chooses research tasks, modifies code and tests the results. According to Baidu, it retains modifications only after they pass separate checks that are hidden from the AI.
Baidu also reported text generation more than six times faster on Kunlunxin P800 chips. China AI Dispatch noted that much of that improvement resulted from activating capabilities the chip configuration had not been using, making it unsuitable for a like-for-like comparison with the GPU gains.
Across three public AI research challenges, Baidu said ERNIE Evolve surpassed the leading public results in two and approached the best result in the other. Its reported NanoGPT Speedrun time was 38.9 seconds, compared with a 39.9-second record. China AI Dispatch said Baidu had not submitted the run to the official leaderboard.
Baidu plans to apply ERNIE Evolve to model training next, followed later by AI-designed tests. The company said expanded self-improvement capabilities should remain subject to people’s ability to understand, oversee and halt the system.
