genai / news / / Decrypt
Baidu's ERNIE 5.1 AI model cost 94% less to train than comparable systems.
ERNIE 5.1 cost 94% less to train than comparable large-scale AI models.
KEY POINTS
- Baidu used 'multi-dimensional elastic pre-training' to shrink ERNIE 5.1's parameters while retaining performance.
- On LMArena, ERNIE 5.1 ranked fourth globally and first among Chinese models for search tasks.
- Baidu implemented a four-stage MOPD pipeline to unify specialist models without skills trade-off.
- ERNIE 5.1 nearly matches leading western models on GPQA and scores 99.6% on AIME26 with tools.
COMPANIES
Summarized by Newsio from Decrypt. How we summarize →