baidu ai search

Baidu World 2026: The Model Built for Baidu Search Is Now an API, and ERNIE 5.5 Flash Preview Promises Apache 2.0 Weights on 25 November

Baidu World 2026 opened in Beijing this morning, and if you were waiting for a big Baidu Search moment you’ll be disappointed. Search itself got no headline launch. The model that has been doing Baidu Search’s heavy reasoning, though, has just become a product anyone can rent.

According to NBD and Yicai, Robin Li launched two “application models” at the opening ceremony: the ERNIE Lightning deep-thinking model (文心闪电深度思考模型) and the ERNIE Yijing micro-expression model for digital humans. Both are live now on Baidu AI Cloud’s Qianfan platform, with APIs open to enterprise customers and partners. The base-model headline was ERNIE 5.5 Flash Preview, which Baidu actually published the day before.

ERNIE Lightning was built for Baidu Search

This is the part SEOs should care about. Zhidx’s live report says ERNIE Lightning “was originally built to serve Baidu Search”. It understands complex requests and calls tools, and later versions will take text, vision and voice input. Baidu credits two in-house techniques: BSA, a sparse attention method based on feedback-driven KV compression, and Looping, a fast/slow thinking scheme. It claims a 30% cut in inference engine cost and an ROI on its own Kunlun chips more than 50% above comparable chips.

One thing to watch: Zhidx gives the parameter count as both 80bn and 800bn in the same article, so I’d treat that figure as unconfirmed until Baidu publishes a model card.

Why does it matter? Taken together with the ¥4.9 Search AI API launched on 30 October, Baidu is now selling both halves of its AI search stack to third parties: the index (as an API) and the reasoning model that decides what to do with a query. If Chinese agents and enterprise apps build on both, Baidu’s view of what counts as a good answer travels a long way beyond baidu.com. That makes visibility in Baidu’s index more valuable, not less, even while the Baidu app itself pushes users towards task completion.

ERNIE 5.5 Flash Preview: priced to undercut, Apache 2.0 promised

Baidu’s release page describes ERNIE 5.5 Flash Preview as the first post-trained model built on the ERNIE 5.5 Flash Base weights released last week. It has 280bn total parameters, 14bn active per token and a 1m-token context. Baidu calls it “an intermediate checkpoint from post-training”, so it’s openly telling us this isn’t the finished model.

Xinhua has the pricing: ¥2 per million input tokens at peak and ¥1 off-peak, ¥8 and ¥4 for output, with a limited-time 50% discount on top. It says the model generates more than 200 tokens a second, and that the weights go to the global open-source community on 25 November.

Here is the twist for readers in London, Paris or Seoul. Baidu’s own page says the Preview weights “will be open-sourced under Apache 2.0 soon”. The Base licence we covered on 9 November excludes the UK, the EU and South Korea. If the post-trained weights really ship under plain Apache 2.0, that geographic carve-out goes away for the model most people will actually use. I’ll believe it when I see the licence file on 25 November.

A token exchange and agents that cite sources

The other launches are less about search but still worth a note. Baidu formally opened Ciyuansuo (词元所), a marketplace for model tokens. Ifeng’s write-up says it uses model fingerprinting and ongoing spot checks to verify suppliers and give them credit ratings. In its first month it signed up more than 100 supply sources and averages over 50m calls and 700bn tokens a day. Baidu also opened the harness engine and Agent API for its office agent Dazi. And it says Kuku, the Wenku/Netdisk creation tool, writes reports that cite professional data and label their sources, drawing on 1.8bn pieces of professional content and 750m Chinese and English papers.

My take

Baidu World 2026 was mostly about enterprise agents, chips (Kunlun M300 next year, N-architecture chips in 2028) and Li’s “intelligent organisation” theme. For search marketers, the real news is that Baidu has turned its search reasoning model into a commercial API at the same moment it’s undercutting rivals on search-grounding prices. Expect more Chinese AI products to be quietly “powered by Baidu” in 2027. Your Baidu index presence and Baidu Webmaster data are worth keeping an eye on for that reason alone.

We’ve added ERNIE 5.5 Flash Base, ERNIE 5.5 Flash Preview and ERNIE Lightning to our ERNIE model updates page.

Dan Taylor is the founder and editor of International Search Insights (ISI), the publication behind ReadIsi, and a Partner and Head of Innovation (Search & AI) at SALT.agency. He has worked in search for more than a decade, in-house and agency-side, and is regarded as one of the West’s leading voices on search engines beyond Google, including Yandex, Baidu, Naver and Cốc Cốc. On ReadIsi, Dan covers the engines and AI answer platforms that English-language coverage tends to miss: Baidu and ERNIE, Yandex and Alice AI, Naver, Cốc Cốc, Shenma, Sogou, WeChat and Weibo search, plus DeepSeek and Alibaba’s Qwen. He also keeps dated records of documented changes, such as the Yandex, Baidu and Naver algorithm update histories and the DeepSeek, ERNIE and Qwen model update histories. His approach is to take learnings from global search engines and apply logic with a high evidence threshold. A claim needs a source, patterns seen in one market are treated as a hypothesis until they are checked in another, and where the evidence is thin he says so. For brands, that means advice built on what can be shown, with measurement that separates being cited from being chosen. Dan is credited with coining the term “edge SEO” and won the inaugural TechSEO Boost research competition in 2018 for his work on CDN and edge changes for SEO. In 2026 he was credited with coining “Infinite Tail” to describe how keyword research is changing with AI search. He is the author of AI Search Marketing For The Travel Sector, writes for Search Engine Journal, Search Engine Land and Travel Marketing Guy, and has advised technology, SaaS, travel and enterprise ecommerce brands, including Cloudflare, GitLab and Airbnb. He has given multiple talks and panels at the International Search Summit in Barcelona, and has also spoken at Optimization Moscow, brightonSEO, Google Search Central Live Deep Dive Asia in Bangkok, Athens SEO and the Asia SEO Virtual Summit 2026, where his talk looked at how Baidu, Naver, DeepSeek, Tencent, Cốc Cốc and Yandex built their own AI answers.