Baidu World 2026 Launches Two ERNIE Application Models: a "Lightning Deep Thinking" Model for Low-Cost 1M-Context Reasoning and Wizshot Micro-Expression for Per-Second 1080p Talking-Head Video

Background and pain points

Baidu World 2026 opened on November 12 with Robin Li describing a two-track strategy: keep building next-generation foundation models, and tune models specifically for agent applications. The two "application models" launched that day come from the second track. Long-document retrieval and long agent conversations suffer when KV cache and cost balloon with context length, and talking-head video needs to be cheap and reliable at volume.

Architecture highlights

The matching text model in the Qianfan docs is ERNIE-Light-DT-1.0 ("Light" maps to Lightning and DT to deep thinking, although Baidu's announcement does not spell out the mapping). Baidu's official OpenRouter description says it is built on open-source GLM models, adds research from Baidu Search's Lightning model, and uses Baidu's own BSA sparse attention with feedback-driven history compression to improve how history is stored and indexed. That cuts KV cache and compute costs and enables a 1,024,000-token context. In other words, this is adaptation and inference optimization on top of GLM, not a new base model trained from scratch. ERNIE-Wizshot-ME is an image-to-video model that makes the person in a photo speak naturally.

Benchmarks

Baidu has published no benchmarks, and third parties such as Artificial Analysis have not scored it yet. What can be confirmed are the official specs and prices: on Qianfan, input costs 0.008 yuan per 1K tokens and output 0.028 yuan per 1K, roughly 4x and 3.5x the peak list price of ERNIE 5.5 Flash Preview. OpenRouter shows about 77 tok/s throughput and about 0.9 s P50 latency (live third-party data). Wizshot-ME lists at 0.7 yuan per second, so a 15-second clip costs about 10.5 yuan.

Getting started

In China, call the Qianfan v2 chat/completions endpoint with model ernie-light-dt-1.0, toggle thinking with thinking, and set depth with reasoning_effort (low, high or max; max is the default). For latency-sensitive classification or extraction, lower the effort or turn thinking off. Outside China, use baidu/ernie-light-dt-1.0-intl on OpenRouter, which supports tool calling and JSON Schema structured outputs. With no public benchmarks, run an A/B on your own long-document samples against ERNIE 5.5 Flash Preview before going to production.