NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Baidu's New Model Holds 280 Billion Parameters but Wakes Only 14 Billion

Baidu released ERNIE 5.5 Flash Preview, a 280B-parameter model that activates 14B per token, and says its weights will be open under Apache 2.0 by November 25.

2026-11-21 · 1037 words · NeuroAI
Baidu's New Model Holds 280 Billion Parameters but Wakes Only 14 Billion

A game studio had a familiar problem: the model could design a whole level, but nobody could afford to keep it running through an eleven-hour coding session. Baidu says it has a way to make only a small fraction of the model show up for each token.

The model is ERNIE 5.5 Flash Preview (文心5.5 Flash Preview), released on November 11, 2026. The catch, for developers who want to inspect it rather than call it through an API, is that the weights are not out yet. Baidu has promised them by November 25.

What Baidu actually released

The official release page describes ERNIE 5.5 Flash Preview as the first post-trained model built on ERNIE 5.5 Flash Base, which Baidu had put out earlier. The headline numbers are 280 billion total parameters, 14 billion activated per token, and a context window of one million tokens. Baidu describes a sparse-attention architecture and a multi-expert design that keeps active compute low.

The API is live on Qianfan (千帆), Baidu's cloud model platform. Baidu's product page says the model generates more than 200 tokens per second.

The release landed one day before Baidu World 2026 (百度世界2026), where chief executive Robin Li used his keynote on November 12 to argue that the competition was shifting from whether a model is strong to whether an organisation can absorb it. The model was the concrete proof point on the stage; the management thesis was the wrapping.

Why 14 billion matters more than 280 billion

A parameter count is a measure of capacity, not of cost. A dense model with 280 billion parameters would fire all of them for every token, which makes long, agentic tasks expensive in both time and memory bandwidth.

A sparse mixture-of-experts (MoE) model stores the capacity but routes each token to a small subset of it. Fourteen billion active parameters out of 280 billion is roughly a twentieth of the whole. That ratio is the commercial argument: a large model (大模型) that behaves like a much smaller one on the invoice.

This is not new in principle, but the scale is. A one-million-token context on a model that activates only 14 billion parameters per token is aimed at a specific workload: long-horizon agents that read large codebases, hold many documents in mind, and keep working for hours. Short chat prompts do not need a million tokens. Coding agents and document pipelines do.

Baidu's own demonstration leans on exactly that. The company says the model completed an eleven-hour task in Unreal Engine 5.8, merging assets from several generations of the engine into one project and writing a third-person wuxia game in C++. Because it is a text-only model and cannot see the screen, the system turned playtesting faults — a character hovering, a running direction that flipped, a camera in the wrong place — into measurable code tests, then iterated against a growing test suite.

That story is more revealing than a benchmark table. It describes a model being judged on whether it can close a loop by itself, not on whether it can answer a question once.

Beyond the game demo

Baidu also showed 3D modelling work built through Blender and CAD tooling, and said the model had been trained on finance and legal document workflows so it could deliver spreadsheets, presentations and reports rather than prose about them. Video creation was the fourth demo track: the model gathered and edited material, adjusted pacing, and produced a finished cut.

The common thread is tool use. Baidu is not selling ERNIE 5.5 Flash Preview as a chatbot that talks well. It is selling it as a component inside an agent harness that keeps calling tools until a task is done. The model's own release notes credit external art and asset packs in many of the demos, a small but useful admission that the system assembles more than it originates.

The open-source promise

Baidu says the weights will be released under Apache 2.0, a permissive licence, by November 25, and that the model can be deployed with both SGLang and vLLM. An earlier version, ERNIE 5.5 Flash Base, had been released earlier in the month.

If that date holds, developers will be able to check the parameter claims themselves, quantise the model, and run it on their own hardware. If it slips, the preview stays what it is today: an API with an attractive active-parameter ratio and a set of company-produced demos.

The distinction matters for a company that has spent years being judged on whether its open releases are real. A permissive licence on a 280-billion-parameter sparse model would be a substantive contribution to the open-weight ecosystem. A delayed promise would simply be another preview.

Honest limitations

Every specification above comes from Baidu's own release page, product page or its presentation at Baidu World 2026, as reported by Chinese state and financial media. I found no independent benchmark of the "more than 200 tokens per second" claim and no third-party reproduction of the eleven-hour game build. The generation speed will depend heavily on hardware and batch size.

The active-versus-total parameter split is a company disclosure. It has not been verified by reading the released weights, because at the time of writing the weights had not been released. Benchmark scores circulating in round-up posts vary between sources and are not used here.

I did not cover the other products announced at the same conference, including Baidu's token marketplace, because they are separate developments and are better assessed on their own. Pricing was also left out: the figures published in secondary posts were not traceable to an official price page that I could confirm.

What readers can do now

  • If you are a developer, wait for November 25 before doing capacity planning. The active-parameter ratio is only useful once you can measure latency and memory on your own hardware.
  • If you already call the Qianfan (千帆) API, run your own long-context task rather than trusting the demo. A million-token window is only valuable if retrieval and attention hold up on your documents.
  • If you are tracking open-weight models, watch the licence, not just the release. Apache 2.0 on a model of this size would matter more to downstream products than another benchmark point.

Related coverage

Cite this article

NeuroAI (脑机网), "Baidu's New Model Holds 280 Billion Parameters but Wakes Only 14 Billion," 21 November 2026, https://neuroai.site/a/na-model-ernie55-flash-preview

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories