Baidu announced on 11 November 2026 that API access to ERNIE 5.5 Flash Preview is available on its Qianfan platform. For researchers outside China, the announcement raises a practical question: what does Baidu actually say about access, terms and data handling, and what remains unstated?
What Baidu announced
Baidu’s product-update page for the model is dated 11 November 2026 and states that “the API is now available on the Qianfan platform.” The page also lists the model on OpenRouter. Baidu describes the model as having 280 billion total parameters with 14 billion active, and a 1M-token context.
On capabilities, Baidu says the model is strong at coding, 3D modeling, knowledge work and content creation, and describes long-horizon planning and game-development support across Unreal, Unity and Godot. These are Baidu’s own claims; the page we reviewed was a product announcement and we did not find independent evaluation of them. Baidu itself notes that the model is “an intermediate checkpoint from post-training,” so some capabilities may still fall short.
Relationship to the open base model
Six days earlier, on 5 November 2026, Baidu published a separate product update for ERNIE 5.5 Flash Base, an open-weights base model. The Preview is a different, API-served checkpoint. Baidu’s Preview page says the weights “will be open-sourced under Apache 2.0 soon” but gives no date, so researchers should not treat open weights for the Preview as available yet.
What Baidu does not state
- Pricing. The Baidu page does not state prices. Check the Qianfan console or the listing for the route you plan to use.
- Data residency and handling. The page gives no details on where prompts are processed or stored, retention, or use of inputs for training. It links only to Baidu’s general privacy policy.
- Usage terms. The page gives no service-specific terms for the Preview API.
What this means for research offices
This is an availability notice, not a due-diligence document. Before researchers send any data to a hosted Chinese model, research offices will want to check the provider’s current terms and privacy documentation directly, confirm where data is processed, and apply institutional rules on research security, export controls, human-subject data and unpublished work. Using the same model through a third-party router such as OpenRouter may involve a different set of terms from using Qianfan directly, so each route needs its own review. Reproducibility is a further consideration: a preview checkpoint can change, so record the model identifier and access date in methods notes.
