Baidu has released ERNIE 5.5 Flash Preview, a new large AI model built for long coding and office work, and says it will open-source the model's weights before 25 November. The model, announced on 11 November, is the first post-trained version of the ERNIE 5.5 Flash Base checkpoint Baidu published on 5 November, and it arrives as Chinese AI labs compete to offer capable models that cost less to run.

The headline numbers are striking. ERNIE 5.5 Flash Preview has 280 billion parameters in total, but only 14 billion are active for any single token, a design known as a mixture of experts. It supports a context window of one million tokens, meaning it can take in very long documents or codebases in a single request, and Baidu says it generates more than 200 tokens per second. It is text-only: it takes in and produces words, not images.

What the model is for

Baidu pitches the model for coding, 3D modelling, knowledge work and content creation. The emphasis is on long, multi-step jobs rather than short answers: the kind of work where an AI agent is given a task and left to work through it, checking its own output as it goes.

The company's showcase demo was deliberately ambitious. Baidu says the model wrote a complete third-person wuxia game in C++ for Unreal Engine 5.8 in a single 11-hour run, without being able to see the screen. It turned bugs found in playtesting, such as a floating character, into code tests and fixed them. Other demonstrations included an airport and airliner built in Blender by two agents working together, a 20-page research report and a 116-second animated short in 20 art styles, driven by more than 10,000 lines of Python.

These are Baidu's own demonstrations, chosen to show the model at its best, and they should be read that way. A demo that runs for 11 hours is impressive, but it is also a single run on a task Baidu selected. The real test will be how the model behaves on ordinary work chosen by someone else.

Baidu's own benchmark scores

Baidu has published a set of benchmark results for the model, all from its own testing, with no independent evaluation yet available. On Terminal-Bench 4.0, which tests long tasks in a computer terminal, the model scored 46.5%, ahead of DeepSeek V4.1 Flash and Hy4 on Baidu's figures but well behind Claude Opus 5.5 at 66.4%.

On DeepSWE, which tests fixing bugs in real software repositories, it scored 71.2%, behind DeepSeek, GPT-6.1 Sol, Hy4 and Claude Opus 5.5. On AutomationBench, which covers business workflow automation, it scored 61.9%, slightly ahead of DeepSeek V4.1 Flash. On office work the picture was stronger: 67.9% on OfficeQA Pro V2, which tests questions over office files, and 64.2% on APEX-Agents, which tests long office agent tasks, though it trailed Hy4 and Claude Opus 5.5 on the latter.

The pattern is consistent: competitive on office and automation work, weaker on fixing real code. Baidu itself describes the model as an intermediate checkpoint from post-training, and warns that some capabilities may still fall short until the official release, because training is still under way. Scores may move as that work continues.

Access and open weights

The model's programming interface is live now on Baidu's Qianfan platform and on OpenRouter, priced at $0.28 per million input tokens and $1.13 per million output tokens. Baidu says the weights will be released under the Apache 2.0 open-source licence before 25 November, which would let developers download and run the model themselves. Until those files appear, that part of the announcement is a promise rather than a fact.

The release also underlines how crowded the efficient-model race has become. The bet behind ERNIE 5.5 Flash Preview is that a model with a large total size but a small active footprint can deliver strong results at a low running cost, and that developers will trade a little raw capability for speed and price. Whether that bet pays off will become clearer once independent testers have had time with the model and the weights are actually available.