Baidu released ERNIE 5.5 Flash Preview on November 11. It's the first post-trained model in the ERNIE 5.5 Flash line that you can put straight to work. The API is live now, and Baidu has promised open weights by November 25, so you'll be able to download the model and run it on your own servers.
What happened
Baidu had already released the ERNIE 5.5 Flash Base model. This Preview builds on it with post-training, mainly aimed at getting the model to finish an entire job by itself inside an AI agent framework.
Key numbers:
| Item | Value |
|---|---|
| Total parameters | 280B |
| Parameters active per request | 14B (mixture-of-experts, so only a slice activates) |
| Context length | 1M tokens, enough to load a large codebase or a big stack of files in one pass |
| Model type | Text only, can't look at images directly |
| Where to use it now | Baidu's Qianfan platform, OpenRouter (has a free tier) |
| Open source | Weights public by November 25 |
A large total parameter count with a small active count means the model holds a lot of knowledge but only uses a small piece for each answer, so it runs fast and costs less.
What Baidu shows it doing
Baidu's launch page is full of examples where the agent carries a job from start to finish:
- Game development: In a continuous 11-hour run, it built a third-person wuxia game in Unreal Engine 5.8. Since it can't see the screen, it turned problems like "the character floats in midair" and "the camera is misaligned" into automated code tests, then fixed things round after round until the tests passed. It also made more than a hundred mobile mini-games from design docs.
- 3D modeling: It used Blender's Python API to build an airport and an airplane, and to recreate Japan's Horyu-ji temple. It also drew editable engine and lunar rover models in FreeCAD.
- Office files: It delivered a 20-plus page research report, a personal finance Excel sheet with a Monte Carlo simulation, plus slides and PDFs.
- Video: From research and scripting through voiceover and editing, one instruction produced a science explainer short.
These are hand-picked examples, and the launch page notes which ones used outside assets and other generation tools. Don't expect the same results on your first try.
How to read the benchmarks
The chart comes from Baidu's launch page. Starred scores are each lab's self-reported results. The model ranks first on AutomationBench (workflow automation) and SpreadsheetBench (spreadsheet work), but clearly trails the top closed-source models on Terminal-Bench (command-line tasks) and DeepSWE (real-world code fixes). Models in a similar class include DeepSeek V4.1 Flash and GLM-5.3, which you can also host yourself.
What this means if you use AI agents
If you want to move your AI agent onto a model that's cheap, handles long context, and that you can later run locally, this one is worth watching. Its most practical use is automation and spreadsheet work, where its benchmarks look strongest.
To try it now, have your AI agent switch the model on OpenRouter to baidu/ernie-5.5-flash-preview:free, then compare it on a small spreadsheet or automation job you normally hand to your agent. Once the weights are open, check whether your own machine can handle them.
Watch out
- This is a preview. Baidu says post-training is still underway, so there may be gaps from undertraining, and the final release will change.
- No image input. Screenshots and design files need a separate vision tool.
- Open source isn't confirmed yet. Public release by November 25 is Baidu's promise. Check the license terms when they land to see whether commercial use is allowed.
- Self-hosting is demanding. Even with only part of the 280B parameters active, all the weights have to fit in memory, and a typical laptop won't run it.
Comments (0)
No comments yet. Start the conversation.