Picture a factory floor in the 1800s, the moment the first electric motors replaced the steam engine — the same kind of shift Baidu now promises with ERNIE 5.5 Flash. The machines looked the same, but everything about how the work was organised had to change.
Robin Li, Baidu’s co-founder and chief executive, reached for exactly that image on stage at Baidu World 2026 in Beijing — moments before unveiling ERNIE 5.5 Flash, the company’s new flagship AI model.
“If AI is the new electricity, then the intelligent organization is the new factory of the AI era, where productivity truly happens,” he told the audience, according to PRNewswire. The line was the keynote’s thesis, and everything Baidu announced afterwards was built to prove it.
The centrepiece was ERNIE 5.5 Flash Preview, unveiled on November 12, 2026, alongside a marketplace for AI model services, upgraded agents, new chips, and fresh numbers from Baidu’s driverless car service.
For a company that spent two decades as China’s search giant, the message was clear. Baidu no longer wants to sell you an AI model. It wants to rebuild the organisation around one.
ERNIE 5.5 Flash: 280 billion parameters, built for speed
The headline model, ERNIE 5.5 Flash Preview, is Baidu’s latest foundation model, designed first and foremost for efficient inference. It carries 280 billion total parameters, with 14 billion activated for each token it generates.
The result, Baidu says, is a model that performs strongly in coding, 3D modelling, knowledge work and content creation while decoding extremely fast, with a context window stretching to 1 million tokens.
On benchmarks including Terminal-Bench and AutomationBench, Baidu reports competitive scores, though it has not published a full scorecard. Developers can use the model now through the API on Baidu AI Cloud’s Qianfan platform, and it is also available free on OpenRouter.
The weights themselves will be open-sourced by November 25. That puts ERNIE 5.5 Flash in the same open-weight conversation as Meituan’s LongCat-2.5, which Baidu’s rival released under an MIT licence days earlier.
Pricing is aggressive: input tokens cost 1 renminbi per million, output tokens 4 renminbi per million, and cached input just 0.04 renminbi per million — roughly a tenth of what Western frontier models charge for comparable work.
ERNIE Evolve: the model that improves itself
Behind the speed claims sits ERNIE Evolve, Baidu’s proprietary recursive self-improvement system — a training loop that refines inference efficiency over time, making the same intelligence cheaper to run.
The most expensive frontier models today, from Google’s Gemini 4 Carbon to Anthropic’s Claude Fable 5.2, compete on capability — but Baidu is competing on a different axis: how much intelligence you get for each yuan.
Two application models: video that feels, thinking that loops
Alongside the flagship, Baidu introduced two specialised application models. The first, ERNIE Wizshot Micro-Expression, is built for highly expressive video creation. It processes text, images and speech together, and it can handle long-form text input of up to 4,000 Chinese characters.
Baidu says it leads the industry for podcasts and short animated dramas, at half the typical cost. That is a telling niche: short-form drama is a booming market in China, and a model that can generate expressive performances cheaply is aimed squarely at creators, not researchers.
The second, ERNIE Light Deep Thinking, tackles complex requests and tool calls at what Baidu calls search-scale speed. It pairs a fast-thinking model that serves users immediately with a slow-thinking model that continuously refines the answer in a closed loop.
Token Exchange: a stock market for AI models
The most unusual announcement was not a model at all: Token Exchange, a marketplace that connects AI model providers with customers, letting them compare services side by side with clear pricing and performance data.
Li described it as an exchange for foundation models, where businesses and individuals can find, buy and use trusted model tokens in one place, and quality providers can reach more customers.
Trust is the hard part of any marketplace. Baidu says Token Exchange rests on two pillars: verification and credit rating. Each resource is verified by comparing model fingerprints and testing capability, performance and availability, while providers are rated through regular spot checks and end-to-end monitoring of every call.
The exchange is already running. One month in, it has onboarded more than 100 model resources and now handles a daily average of more than 50 million calls and over 700 billion tokens.
The intelligent organisation, measured in agents
Li’s keynote was not just a product list. He laid out three principles for how organisations should evolve, then showed Baidu’s own numbers as proof.
The first is to break down job boundaries. With AI letting one person act in multiple roles, Li said organisations will be built around tasks rather than jobs.
“Jobs won’t disappear, but they will converge into roughly two types: Builders, who create new things, and Sellers, who connect with people to forge emotional bonds and earn their trust,” he said.
The second is to capture tacit knowledge — the unwritten know-how inside people’s heads — and turn it into a shared company asset. “This is a capability that competitors can’t copy, hire away, or buy. That’s what a real moat looks like,” Li said.
The third is to design closed-loop processes. His formula was simple: measurability plus looping equals continuous evolution, so that “the organization becomes a system that never stops learning and evolving.”
Inside Baidu, the results are striking. Over the past year, the average number of daily active agents inside the company rose from 5,000 to 37,000, an increase of more than 600 percent.
Those agents now complete more than 1 million tasks a week, up from 84,000, and Baidu has built 4,300 workspaces and captured more than 20,000 organisational skills.
Agents get an upgrade: Yijing, DuMate, KOOKO and Famou
Baidu’s agent lineup got a full refresh. Baidu Yijing, the hyper-realistic digital human agent, can now create short animated dramas covering concept, script, storyboard and final cut — turning a single idea into a finished production in one stop, with one-click localisation for global audiences.
Production costs just 0.2 renminbi per second of video.
DuMate, the general-purpose agent, has been upgraded to serve as the unified gateway to Baidu’s full range of specialised agents, and Baidu has fully opened its Harness engine and Agent API.
KOOKO, the AI-powered platform for documents and creative work, was upgraded from GenFlow 3.0. It draws on 1.8 billion professional resources from Baidu Wenku and Baidu Drive, plus 750 million Chinese and English academic papers, to help research, write and create reports, slide decks and websites.
Famou 3.0, the self-evolving agent for complex operational challenges, now serves customers spanning 25 of China’s 41 industrial categories.
Chips and cars: the foundation underneath
None of this runs without hardware. Baidu AI Cloud launched the Tianchi 1024-card Supernode, an all-optical interconnect supernode built for high-speed data transmission between accelerators.
Kunlunxin, Baidu’s AI chip business, announced that the M100 inference chip has shipped as planned, while the flagship M300 is expected to ship in 2027.
“We believe inference will be the primary driver of sustained long-term growth in the AI chip market,” Li said. That bet mirrors the wider industry.
Alibaba’s Qwen4-Max preview, launched the same week, is also built for the inference era — a world where the money is made not by training giant models, but by running them cheaply at enormous scale.
On the road, Apollo Go, Baidu’s autonomous ride-hailing service, now covers 28 cities worldwide and has logged more than 350 million autonomous kilometres, including over 240 million in fully driverless mode.
It has launched fully driverless commercial operations in Abu Dhabi and Dubai, and partnered with CAR Inc. in Hainan to pioneer what Baidu calls the world’s only driverless car rental model.
“Embodied AI is now at a stage comparable to where autonomous driving stood in 2016,” Li said — a reminder that the robots walking factory floors today are roughly where self-driving cars were a decade ago.
What it means next
Baidu’s pitch is a bet that the next phase of AI is not about a single smarter model, but about organisations that run on agents, models traded like commodities, and inference so cheap that intelligence becomes a utility.
Li closed by naming three opportunities he believes intelligent organisations will open: asset-light operations, new competitive moats, and the creation and discovery of new markets. “So now is the moment to drive the intelligent evolution of organizations,” he said.
Whether the vision holds depends on execution — on whether the open weights promised by November 25 actually arrive, and whether Token Exchange can earn the trust it needs to become a real marketplace.
But the scale of the announcement is hard to ignore: a flagship model, a token exchange, four upgraded agents, new chips and a driverless fleet, all in one morning.
“Baidu is ready to work with everyone to rethink, redesign, and recreate the organization,” Li said. “Let us take the initiative in organizational transformation and win the future through intelligent evolution.”
For a company that once defined itself by search, that is a very different future — and, according to PRNewswire’s report of the event, it starts now.
