Felix Laurent
Baidu released ERNIE 5.5 Flash Preview on November 11, 2026. It is the first instruction-tuned model built on ERNIE 5.5 Flash Base, the open checkpoint Baidu published six days earlier. The model has 280 billion parameters, of which 14 billion are active per token, and a 1-million-token context window.
The API is live on Baidu Qianfan and free on OpenRouter for now. Baidu says the weights will be open-sourced under the Apache 2.0 license by November 25. On Qianfan the model generates more than 200 tokens per second for each user.
What ERNIE 5.5 Flash Preview can do
Baidu built the model for long agent tasks and showed it in four areas: game development, 3D modeling, office work and video. Each demo ran end to end inside an agent harness, and Baidu credits the outside assets it used.
The headline demo is a game. In a single 11-hour coding session, the model merged assets made for several versions of Unreal Engine into one Unreal Engine 5.8 project. It then wrote a complete third-person wuxia game in C++, set among the sandstone peaks of Zhangjiajie.
The model is text-only and cannot see the screen. So it turned bugs found in playtests, such as a character hovering above the ground or a misplaced camera, into tests in code. A growing test suite drove each fix. Baidu also says the model built nearly 100 mobile mini-games from design documents in engines such as Cocos and Unity.
Other demos cover the rest of Baidu's list:
- 3D and CAD. The model drove Blender through Python to build an airport, an airliner and the Hōryū-ji temple. In FreeCAD it modeled an editable W16 engine and a lunar rover concept.
- Office work. It wrote 20- and 24-page research reports and a personal finance workbook with Monte Carlo simulation. Baidu trained it on 300,000 workspaces holding 10 million reference files.
- Video. In about one to two hours per video, it handles the script, storyboard, assets and final render. One demo is a 116-second animated short drawn entirely in Python code.
Demos show the best case, so the benchmark table is the better guide.
ERNIE 5.5 Flash Preview benchmarks
Baidu compared the model with other Chinese models and two US frontier models. The list covers Hy4, DeepSeek V4.1 Flash, GLM-5.3 and its Flash version, GPT-6.1 Sol and Claude Opus 5.5. Many of the rival scores come from Baidu's own runs, which the table marks.
The model does best on office and tool work. It leads every model in the table on AutomationBench with 61.9, ahead of Claude Opus 5.5 at 58.0. It also tops SpreadsheetBench V2 at 36.4 and MCP-Atlas at 84.6. On OfficeQA Pro V2 it scores 67.9, behind only GPT-6.1 Sol.
Asterisks mark scores from Baidu's own runs. Where a model has two scores, the table shows the officially reported one.
Hard coding and reasoning are another story. On Terminal-Bench 4.0 the model scores 46.5, against 56.1 for GPT-6.1 Sol and 66.4 for Claude Opus 5.5. On Humanity's Last Exam without tools, it gets 46.4 against 61.4 for Opus. Among open models, it trails DeepSeek V4.1 Flash on DeepSWE, 71.2 to 77.0, and Hy4 on APEX-Agents and BrowseComp.
Those results fit a model with 14 billion active parameters. Its efficiency comes from the architecture it shares with the base model.
How ERNIE 5.5 Flash stays fast
The Preview uses the same architecture and tokenizer as ERNIE 5.5 Flash Base. It has 43 layers. In 37 of them, attention stores one compressed cache entry per 128 tokens. Only six keep full context. That keeps memory use low at 1 million tokens.
A multi-router latent mixture of experts picks 10 of 512 experts for each token, plus one shared expert. The last layer of the network also serves as the draft model for speculative decoding, which speeds up generation. The tokenizer has about 200,000 tokens and splits numbers from right to left in groups of three.
We covered the design in more detail when Baidu released ERNIE 5.5 Flash Base. Baidu also used its ERNIE Evolve automated research system to speed up inference for this model.
Is ERNIE 5.5 Flash Preview worth trying
For agent work in office tools, yes. It is free on OpenRouter during the preview, fast on Qianfan, and strong on spreadsheet and automation tasks. Once the Apache 2.0 weights arrive, teams will be able to run and tune it on their own hardware.
For hard coding and reasoning, frontier models stay well ahead. Many benchmark scores are Baidu's own, and the demos are hand-picked. The open weights, due by November 25, will show how much of this holds up in independent tests.
Sources
- ERNIE 5.5 Flash Preview, Baidu – November 11, 2026
- Baidu ERNIE on X – November 11, 2026
- ERNIE 5.5 Flash Preview on OpenRouter
- ERNIE 5.5 Flash Base, Baidu – November 5, 2026
