BaiduのErnie 5.1、事前学習コストを94%削減
Baiduが「ernie 5.1」をリリース。事前学習コストを94%削減しつつ、中国のAIベンチマークで首位に立つと主張するモデル。
baidu unveiled ernie 5.1, a 94% cost-efficient iteration of its ernie 5.0 model, using an elastic training framework to share weights and reduce parameters. it claims the model surpasses deepseek-v4-pro in agent tasks and rivals google gemini 3.1 pro in reasoning, ranking 1st in china and 4th globally on the arena search leaderboard. technical improvements include a decoupled reinforcement learning infrastructure and a four-stage fine-tuning pipeline to mitigate the 'seesaw effect' where different skills degrade during training. the model is integrated into baidu's applications but remains closed-source, preventing independent verification of these benchmarks.