Baidu's efficient long-context model, built on open GLM weights with Baidu Search's Lightning research and Baidu Sparse Attention. 1M-token context that stays affordable at length: feedback-driven history compression cuts KV-cache storage and compute while keeping quality, so long documents and extended agent histories fit a low per-token price.
▼ERNIE-Light-DT-1.0 is somewhat expensive at $1.13/M tokens input and $3.94/M tokens output. Cached input reads are priced at $0.11/M tokens, making repeated context cheaper. The model supports extended reasoning, a 1M context window.