Baidu has released an API-accessible preview of ERNIE 5.5 Flash, a 280-billion-parameter model with 14 billion parameters active per token and a one-million-token context window. Developers can try it through Baidu’s Qianfan platform now; the promised open weights are not yet available.
Watch Desk analysis
What happened
Baidu released the preview on 11 November, according to PanDaily’s report. It is a post-trained version of ERNIE-5.5-Flash-Base, which Baidu open-sourced on 6 November. The report says Qianfan lists the model endpoint as ernie-5.5-flash-preview, with a default limit of 60 requests and one million tokens per minute.
The weights are a separate matter. Baidu says they will be open-sourced by 25 November, but the report says they were not publicly downloadable when checked. Baidu also describes the preview as an intermediate post-training result and warns that capability gaps may remain. Its claims about benchmarks and fast decoding are not backed by benchmark details in the report.
Why it matters
The distinction is practical: developers can access the model through a hosted API, but cannot yet download and run its weights themselves. The one-million-token context and stated parameter counts give a sense of scale, while the API limits and hosted access define what people can test today. “Open weights by 25 November” is a useful date to watch, not a feature already in hand.
Our read
This is a substantive model release, but the interesting test is not whether the headline numbers look grand. It is whether developers can get useful results from the preview, and whether Baidu follows through with weights and evidence that others can inspect. For now, there is something to try, but not yet a public model to install.
What to watch
- Whether Baidu publishes the promised weights by 25 November.
- Whether independent evaluations clarify the preview’s performance and remaining capability gaps.
- How Qianfan access limits and costs affect practical testing.
Discussion spark: Would you test a model through a hosted API while waiting for its weights, or is local access important enough that you would wait?
Sources and evidence
- Baidu Releases ERNIE 5.5 Flash Preview on Qianfan, With Open Weights Promised by November 25 (11 November 2026, 07:10 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.