Baidu opens ERNIE 5.5 Flash Preview API ahead of planned weight release

Baidu announced ERNIE 5.5 Flash Preview on November 11 and said it was available through its Qianfan cloud API. Developers can connect the preview to their software through Qianfan while they wait for downloadable weights. Baidu’s Chinese announcement puts the planned release of the weights before November 25. Its English announcement names Apache 2.0 as the intended license.

The preview is the first post-trained version of the ERNIE 5.5 Flash Base model that Baidu listed on November 5. The base model already has FP8 and BF16 downloads under a separate Baidu license. Its model card documents serving that model through a text-completion endpoint. Baidu’s Hugging Face collection still listed those two base checkpoints when checked on November 11.

Qianfan lists separate quotas for generation and reasoning

Baidu says the preview has 280 billion parameters, with 14 billion active when processing a token. The model selects a subset of its learned parameters for each piece of text it processes. Baidu lists a context window of one million tokens, the amount of text the model can take into account in a task.

Qianfan’s November 11 model list puts the same identifier, ernie-5.5-flash-preview, under both text generation and deep reasoning. It gives text generation a default rate limit of 300 requests and three million tokens per minute. The deep-reasoning entry lists 60 requests and one million tokens per minute. Both entries list a one-million-token context window and give one million as the individual input and output limits. The per-minute figures describe the default throughput limits across calls. The platform says any reasoning text it returns before a final answer counts toward the documented maximum output.

Qianfan prices vary by time and cache use

Qianfan bills online inference by token usage. In its November 10 pricing table, the promotional peak rates work out to 1 yuan per million input tokens, 0.10 yuan per million cached input tokens and 4 yuan per million output tokens. Off-peak promotional rates are 0.50 yuan, 0.05 yuan and 2 yuan, respectively. The table lists the same prices for general text generation and deep reasoning.

The listed standard peak rates are 2 yuan for input, 0.20 yuan for cached input and 8 yuan for output per million tokens. Off-peak standard rates are 1 yuan, 0.10 yuan and 4 yuan. Qianfan labels 08:00 to 22:00 as peak hours and 22:00 to 08:00 the next day as off-peak. The pricing entry leaves the time zone and the promotion’s end date unspecified. The table leaves the batch inference price blank for this model.

Baidu’s game and CAD demos used external tools

In Baidu’s 11-hour game demo, the model wrote a third-person game in C++ for Unreal Engine 5.8 using existing assets. Baidu says it turned movement and camera problems found in playtesting into tests it could run in code.

In another demo, Baidu says the model wrote Python code that produced editable engine parts in FreeCAD. Both demonstrations used an agent environment with external tools. Qianfan lists the hosted model under text generation and reasoning.

Baidu’s comparison table gives ERNIE 61.9 on AutomationBench, which tests workplace automation. Baidu reports a score of 58.0 for Claude Opus 5.5 on the same benchmark. The table lists 46.5 for ERNIE and 63.6 for Baidu’s evaluation of Claude Opus 5.5 on Terminal-Bench 4.0. The notes say the models were evaluated at their highest available reasoning settings. The software environment varies by benchmark, and some comparison results come from external reports or public leaderboards.

Baidu describes the preview as an intermediate version in post-training. The company says training is still underway and it will use feedback from real-world use to improve the final version.