Baidu released ERNIE 5.5 Flash Preview on November 11, making the post-trained model's API available on Baidu Qianfan while pledging to open-source the weights by November 251,2.
The model is the first post-trained version built on ERNIE 5.5 Flash Base, the raw pretraining checkpoint Baidu open-sourced on November 6[2]. It carries 280 billion total parameters with 14 billion active per token and supports a context window of 1 million tokens. Baidu says it performs well in programming, 3D modeling, knowledge work, and content creation, and reports competitive benchmark scores with fast decoding speeds. Those are Baidu's own claims; independent benchmark tables were not available in the source material.
The Baidu Qianfan platform's model list, updated on November 10, shows an endpoint named ernie-5.5-flash-preview with a 1-million-token context and a default limit of 60 requests and 1 million tokens per minute. Baidu describes the Preview release as an intermediate result of post-training and warns that insufficient training may still cause capability gaps.
The open-weight commitment remains a promise rather than a delivered artifact. When checked on November 11, the Preview repository on Hugging Face returned an access error, and Baidu's public Hugging Face listing showed only the two Base checkpoints. Baidu's vLLM-ERNIE repository does list Preview in FP4, FP8, and BF16 in its supported-models table, and a deployment guide from the SGLang project describes the chat version as coming soon.
ANALYSIS The five-day gap between the Base checkpoint release on November 6 and the Preview API on November 11 marks a rapid post-training turnaround, though Baidu's own caveat about capability gaps frames this as an early-access artifact rather than a finished product. Whether the November 25 weight release holds will be the next concrete test of Baidu's open-model cadence.