Baidu launches ERNIE Lightning reasoning model and ERNIE-Wizshot-ME video model on Qianfan

Baidu released two application-focused ERNIE models at Baidu World 2026 on November 12: ERNIE Lightning, a deep-thinking model, and ERNIE-Wizshot-ME, a micro-expression video model. Both are live on Baidu AI Cloud's Qianfan platform, with APIs open to enterprise customers and partners. Jia Lei, head of Baidu's application model R&D department, presented them.

ERNIE Lightning

According to Baidu, ERNIE Lightning was first built for Baidu Search. As search moves from retrieval toward answers, agents and multi-step tasks, the model must understand complex requests, reason and call tools, while staying fast for very large numbers of users.

Baidu describes two main technical choices:

• BSA (Baidu Sparse Attention): Baidu says it reworked the standard Transformer with this sparse attention method, which is based on feedback-driven KV-cache compression.
• Looping scheme: a fast-thinking model handles live requests, while a slow-thinking model acts as a judge and feeds back improvements.

The model offers thinking and non-thinking modes and covers coding, logical reasoning, agent and text-processing tasks. Baidu says that with Kunlunxin co-optimization it runs fast inference on a single eight-card Kunlunxin server and cuts inference engine cost by more than 30%.

Qianfan and OpenRouter also list ERNIE-Light-DT-1.0, which Baidu describes as built on open-source GLM models with findings from Baidu Search's Lightning model. Baidu has not said whether that is the same product as ERNIE Lightning.

ERNIE-Wizshot-ME

ERNIE-Wizshot-ME targets emotional expression in AI-generated video. Baidu says it models text, vision and speech together and accepts scripts of up to 4,000 characters. It uses a fine-grained description system of six dimensions and 200 sub-dimensions to turn emotional cues into facial expressions, body movement and mood changes.

A passage-level dubbing feature voices a whole script in one pass instead of stitching roughly 30-second segments, which Baidu says keeps tone and pacing consistent.

Qianfan's product page says the model animates a person in a still image to speak, producing four- to 15-second 1080p videos at a list price of 0.7 yuan per second. Baidu claims video production costs 50% below the industry level, and says a fast version generates one minute of video in about one minute.

Key details

Item	ERNIE Lightning	ERNIE-Wizshot-ME
Type	Deep-thinking (reasoning) model	Micro-expression video model
Availability	Qianfan, API for enterprise customers and partners	Qianfan, API for enterprise customers and partners
Pricing	Not yet disclosed	0.7 yuan per second (list price)
Output	Text, with thinking and non-thinking modes	1080p video, four to 15 seconds
Benchmarks, parameters, context length	Not published	Not published

What this means

Both releases are framed around deployment rather than leaderboard results. For ERNIE Lightning, Baidu's stated priorities are serving latency and inference cost at search scale, supported by its claim of a more than 30% cut in inference engine cost on Kunlunxin hardware. Those figures are Baidu's own, and no independent benchmarks have been published, so its reasoning quality cannot yet be compared with other models.

For ERNIE-Wizshot-ME, the published per-second list price and the 1080p, four- to 15-second output spec give developers concrete numbers to plan around. The 50% cost claim and the one-minute generation time for the fast version are Baidu's assertions and have not been independently verified.

The unresolved relationship between ERNIE Lightning and ERNIE-Light-DT-1.0 matters for anyone selecting a model on Qianfan or OpenRouter. Baidu describes the latter as built on open-source GLM models, and has not said whether the two are the same product. Until Baidu clarifies, teams should not assume the two listings behave identically.