Image: Baidu
Baidu released ERNIE 5.5 Flash Preview, the first model tuned on top of the ERNIE 5.5 Flash Base it published six days ago. It has 280 billion parameters in total, switches on 14 billion of them for each token, and takes a 1 million token context. Baidu says it generates more than 200 tokens per second.
The model is on Baidu's Qianfan API now. Baidu says the weights will be open-sourced under Apache 2.0, and its post puts the release by November 25. That licence is the difference from the Base model, which shipped earlier this month under Baidu's own agreement.
ERNIE 5.5 Flash Preview just landed! 🔥
— Baidu Inc. (@Baidu_Inc) November 11, 2026
Built on a new architecture designed for efficient inference, the latest open-weight foundation model in the ERNIE family delivers stronger performance in coding, 3D modeling, knowledge work and content creation. It keeps decoding fast while posting competitive scores on Terminal-Bench, AutomationBench, and other benchmarks. Here’s the rundown:
- 280B total parameters, 14B active per token
- 1M-token context window
- Generation speed above 200 tokens/sec
Model weights will be released by November 25.
Running a fraction of the model at a time is what keeps a model this size affordable. ERNIE 5.5 Flash sends each token to a small set of specialist sub-networks rather than the whole model, a design called a mixture of experts. We explained how that works in our post on why large models use smaller experts.
Baidu's benchmark chart is less flattering than the launch language. The preview leads two of the eight tests, AutomationBench at 61.9 against 58.0 for Claude Opus 5.5, and SpreadsheetBench V2 at 36.4 against 35.1. It comes second on OfficeQA Pro V2, at 67.9 to 70.0 for GPT-6.1 Sol.
It trails further on the coding tests. Claude Opus 5.5 scores 63.6 on Terminal-Bench 4.0 to its 46.5, and DeepSeek V4.1 Flash reaches 77.0 on DeepSWE v1.1 to its 71.2. On Agents' Last Exam it is fifth at 29.0, behind DeepSeek at 35.2. Some competitor scores carry an asterisk for self-evaluated, and every number in the chart comes from Baidu. No third party has reproduced them.
Image: Baidu
Baidu is also direct about the model's state. Its release page calls the preview an intermediate checkpoint from post-training and warns that some capabilities may still fall short because training is not finished. That fits the scores. It is competitive with models of its size, not ahead of the frontier.
The demos point the same way. Baidu says the model built a third person game in Unreal Engine 5.8 in a single 11 hour coding task, and about 100 mobile mini-games from design documents. The model is text only, so it turned playtest problems into tests in code instead of looking at the screen. Those runs happened inside Baidu's own agent setup, with external art and sound assets supplied.
A free tier is listed on OpenRouter. For anyone deciding whether to try it, the practical case is speed and context, with 200 tokens a second and a 1 million token window on a model that is large on paper and cheap in the part that runs.
The weights are the part to watch. If the Apache 2.0 promise holds when they land, ERNIE 5.5 Flash becomes one of the few large Chinese models anyone can download without a custom licence. Baidu has not said whether the preview's benchmark claims will change once training finishes.
Comments