Baidu’s Free New Model Is Climbing a Coding Leaderboard. Inside Baidu, 37,000 SI Agents Already Work

Baidu walked into its biggest event of the year with two things to show. One was a new model it plans to give away. The other was a set of numbers about how hard the company has been leaning on its own software agents to do everyday work.

Both came out around Baidu World 2026 in Beijing, where founder and CEO Robin Li took the stage. The pitch, boiled down: China’s search giant wants to be known for cheap, open models that do real jobs, and it says it is already running its own business that way.

A big model that only uses a slice of itself

The model is ERNIE 5.5 Flash Preview, released a day ahead of the keynote. On paper it has 280 billion parameters, the internal settings a model learns during training. But it only switches on 14 billion of them for each word-piece it reads or writes.

That design is called a mixture of experts. Think of a hospital with hundreds of specialists where each patient only sees the few doctors they need. You get the knowledge of a very large model with the running cost of a much smaller one. Baidu says the model can take in up to a million tokens at once, enough for a long book or a big codebase, and its release page says it is aimed at coding, 3D modeling, office work and content creation.

Outside testers have started to weigh in. Arena, the site where users vote blind between answers from competing models, said Wednesday that the model was trending toward the No. 2 spot among open models on its WebDev coding board, at 1,668 points. That edged past DeepSeek-V4.1-Flash at 1,664 and Kimi K3 at 1,658. Baidu passed the news along on its own account.

A few caveats belong next to that score. Arena rankings move as votes come in, and “trending toward” is not the same as a settled rank. Baidu itself calls this release “an intermediate checkpoint from post-training” and warns that “some capabilities may still fall short” until the final version ships.

Free weights, cheap tokens

The bigger hook for developers is the price. Baidu says the weights will be released under Apache 2.0, one of the most permissive open licenses there is, so companies can download the model and run it on their own machines. Chinese tech outlet Kuai Keji, as carried by Sina, reported the weights are planned for Nov. 25.

Until then, the model is available through Baidu’s Qianfan cloud platform. According to that same report, the price is 2 yuan per million input tokens at peak hours and 1 yuan off-peak, with output at 8 yuan peak and 4 off-peak, plus a limited-time half-off deal. Baidu also claims generation speeds above 200 tokens per second. Those are the company’s own numbers, not independent measurements.

Baidu says it trained the model hard on long, multi-step jobs. Its examples include building a playable game from a prompt, modeling an airport and airplanes in Blender, and assembling short videos from script to final cut. According to the Kuai Keji report, the video training alone covered more than 150,000 creative setups, from science explainers to animated shorts.

That puts Baidu squarely in the price war that has defined Chinese models this year. DeepSeek, Moonshot’s Kimi and others have pushed open weights and rock-bottom pricing to win developers around the world. A free, efficient model from Baidu adds one more option for any company that would rather run its own copy than rent one from an American lab.

Robin Li’s real argument: fix the company, not just the tools

Li spent his keynote on a different question. If businesses keep spending more on these tools, why aren’t their profits rising to match?

His answer, as Baidu summarized it, is that companies “have replaced the motor without redesigning the factory.” Swapping in a smarter tool is the easy part, he argued. The payoff comes from rebuilding the work itself around agents, in sales, customer service, research and operations.

To make the case, Baidu pointed at itself. The company says that a year ago it set out to rebuild its own operations around this kind of software, and it posted the results.

By Baidu’s count, the number of agents active on an average day inside the company went from about 5,000 to 37,000. Tasks those agents finished went from 84,000 a week to more than a million. Staff built 4,300 shared workspaces and more than 20,000 reusable “skills,” basically saved know-how that any team can pick up. Again, these are internal figures Baidu reported about itself, and it did not say how it counts a “task.”

Li also offered three rules for what he calls an intelligent organization: break down the walls between job roles, turn the knowledge that lives in people’s heads into shared assets, and build processes that feed results back so the company keeps learning. Looking further out, he said he expects most jobs to fall into two buckets: builders who make new things, and sellers who earn trust and keep relationships.

Baidu also used the event to announce two application models on Qianfan, with APIs open to enterprise customers and partners, Every Day Economic News reported. One is a fast “deep thinking” reasoning model. The other is built to read micro-expressions, the small, quick changes on a person’s face. Li said Baidu will keep pushing two tracks at once: frontier models that raise the ceiling, and models tuned for the products people actually use.

For developers, the date to circle is Nov. 25. If the weights land as promised and the early coding scores hold up, Baidu will have one of the strongest models anyone can download for free.