News

Baidu releases ERNIE 5.5 Flash Preview through Qianfan API

Baidu has listed ERNIE 5.5 Flash Preview for use through Qianfan, while the post-trained model's promised Apache-2.0 weights remain unavailable.

D
Nov 19, 2026 · 2 min read

Baidu released ERNIE 5.5 Flash Preview through its Qianfan API on November 19. The unfinished post-training checkpoint is listed for use, while its promised open weights remain unavailable. The company says the model has 280 billion total parameters, activates 14 billion parameters per token and supports a one-million-token context window.

Baidu’s Qianfan launch page and model-specific API documentation list the Preview as available through the platform. The opened evidence does not include an independently completed inference request.

The open-weight release, by contrast, remains a future commitment. Baidu said it plans to publish the post-trained Preview weights under the Apache 2.0 license “soon,” but did not give a date.

Baidu describes the model as a sparse mixture-of-experts system, which routes each token through only part of the network instead of using all 280 billion parameters for every token. According to the company, the underlying ERNIE 5.5 Flash Base architecture has 43 Transformer layers. Its mixture-of-experts layers contain 512 routed experts, with 10 selected for each token alongside one shared expert.

The company says the Base checkpoint was pretrained from scratch on 33.5 trillion tokens. Its training context was extended to 128,000 tokens, while the one-million-token figure applies to fine-tuning and inference. Baidu also says most attention layers compress each 128-token block into one cached entry, while one layer in seven retains token-level full-context access.

Baidu calls the Preview an intermediate checkpoint whose post-training is unfinished and warns that some capabilities may fall short of the eventual release. The existing Base weights are a separate, pretraining-only checkpoint governed by Baidu’s ERNIE 5.5 Flash Base Model License Agreement v1.0, not the promised Apache-2.0 Preview weights.

On Qianfan’s model launch page, Baidu says the Preview generates more than 200 tokens per second, uses FP4 expert weights and a draft head, and defaults to an output length of 64,000 tokens. Those performance statements were not independently measured. Baidu also presented the benchmark results and demonstrations involving game development, 3D modeling, knowledge work and video creation; none was independently reproduced in the opened evidence.

The Qianfan API documentation says the Preview does not support temperature or top-p controls, does not let users disable thinking or adjust reasoning effort, and did not support web search when the documentation was observed.

More news