Baidu's ERNIE 5.5 Flash Preview: 280 billion parameters, 14 billion active, a wuxia game coded in one 11-hour run, and open weights promised by Nov. 25

The backstory: total parameters vs. active parameters

A model's parameters are, roughly, the knowledge and skills stored inside it. More parameters usually means a smarter model, but also more computation, time, and cost every time it answers. A popular design called mixture of experts (MoE) splits a model into many specialist groups and wakes up only the few most relevant to each word it processes. That lets a model be very large while keeping each step relatively cheap.

Post-training is what happens after a base model has read a huge amount of text: it is taught to follow instructions, write code, use tools, and behave the way people want. If the base model is a new graduate, post-training is on-the-job training.

What Baidu released

On November 11, Baidu released ERNIE 5.5 Flash Preview. According to Baidu, it has 280 billion parameters in total but activates only 14 billion for each token, and it can take in about 1 million tokens at once. A token is the basic unit of text a model reads; a million tokens is several books or a mid-sized codebase. Baidu's official ERNIE account on X said users get more than 200 tokens per second on Baidu's Qianfan cloud platform.

It is the first post-trained version of ERNIE 5.5 Flash Base, a base model Baidu open-sourced earlier. The API is live on Qianfan, and Baidu's page also links to OpenRouter. Baidu says the weights — the model files themselves — will be released by November 25 under Apache 2.0, a permissive license that lets individuals and companies use and modify the model for free, including in commercial products.

The showcase is eye-catching. Baidu says that in a single 11-hour coding task, the model merged assets built for several generations of Unreal Engine into one UE 5.8 project and wrote an entire third-person martial-arts game in C++, set among the sandstone peaks of Zhangjiajie. Because it is a text-only model and can't see the screen, it turned playtest problems — a character hovering above the ground, an inconsistent running direction, a misplaced camera — into measurable tests in code and kept iterating until each goal was met. Baidu also says the model built nearly 100 mobile mini-games from design documents, modeled an airport and an airliner in Blender by writing code, designed a moving engine and a lunar rover concept in FreeCAD, and produced 20-plus-page research reports on its own.

A caveat worth stating plainly: these demos and benchmark scores come from Baidu. The model ran inside an agent harness and used third-party assets and tools, and no one has independently reproduced the results yet. Baidu itself says the Preview is an intermediate checkpoint from ongoing post-training and may have weaknesses from incomplete training.

Why it matters

First, openness. If Baidu delivers Apache 2.0 weights for a 280-billion-parameter model aimed at coding and agents, companies and developers can run and customize it on their own servers without sending data to someone else's cloud. That raises the pressure in the already crowded open-model race.

Second, long-running autonomous work. AI coding used to mean finishing a few lines or fixing a function. Now model makers compete on whether a model can plan, test, and rework a complete project for hours on its own — a key step in AI moving from assistant to coworker.

What it means for you

You may never use this model directly, but the apps, mini-games, and office tools you use could increasingly run on open models like it, which helps companies cut costs. For anyone learning to code, making indie games, or doing 3D design, the barrier keeps dropping. Just keep in mind that company demos show the best cases; real-world use can still go wrong, so check important work yourself.

Original sources published: IT Home, Nov 11, 2026, 14:35 Beijing time (UTC+8); Baidu ERNIE official X account (@Baidu_ERNIE), Nov 11, 2026, 15:56 Beijing time