zhineng.dev
English
Get started
← All models

ERNIE-Light-DT-1.0

Published Publisher Baidu

Overview

ERNIE-Light-DT-1.0 is a lightweight, long-context text-to-text model available through Baidu AI Cloud’s Qianfan platform. Baidu’s November 12, 2026 announcement describes ERNIE Light Deep Thinking as an application model originally developed for multimedia search and designed for complex requests and tool calls. (prnewswire.com)

Capabilities and uses

  • Uses Baidu Sparse Attention (BSA) and feedback-based history compression to reduce historical-information storage and computation costs. Thinking mode can be enabled or disabled. (cloud.baidu.com)
  • Baidu lists long-document question answering and summarization, multi-turn conversations, agent tool calls, batch text processing, and cost-sensitive high-frequency calls as use cases. (cloud.baidu.com)
  • Baidu’s international model list specifies a 1M-token context window and a maximum output of 128K tokens. Its listed default rate limits are 60 requests per minute and 1,000,000 tokens per minute. (intl.cloud.baidu.com)

Limitations

The listed model accepts text and returns text; the cited product page does not establish direct image, audio, or video input support. Baidu’s international documentation says ERNIE thinking models do not support thinking_budget, and max_tokens limits reasoning and final-answer content combined. Outputs should be checked for accuracy. (cloud.baidu.com)

Pricing

Baidu Qianfan China, pay-as-you-go: CNY 8 per million input tokens, CNY 0.8 per million cache-hit input tokens, and CNY 28 per million output tokens. Baidu states that these rates are uniform throughout the day, without peak/off-peak tiers. The promotion shown elsewhere on the same product page applies to ERNIE-5.5-Flash-Preview, not this model. Source: Baidu AI Cloud product and pricing page; checked November 14, 2026. (cloud.baidu.com)

Baidu’s international pricing page was also checked on November 14, 2026; an official international price for this model could not be verified there. (intl.cloud.baidu.com)

Release and official sources