Baidu Releases ERNIE 5.5 Flash Preview on Qianfan

Baidu has released ERNIE 5.5 Flash Preview, the first post-trained model built on the ERNIE 5.5 Flash Base it published earlier, and has put the model on its Qianfan API. In a product update dated 11 November 2026, Baidu said the model “has 280B total parameters with 14B active and a 1M-token context,” and that it is aimed at coding, 3D modelling, knowledge work and content creation.

The international Qianfan model list already carries the endpoint. The model ID is ernie-5.5-flash-preview. Context length, maximum input and maximum output are each listed as 1 million tokens. The default rate limits are 60 requests per minute and 1,000,000 tokens per minute.

Baidu says the weights “will be open-sourced under Apache 2.0 soon.” That is a plan, not a date. The 11 November pages give no calendar day for the files, and none is attached here.

ERNIE 5.5 Flash Preview is an intermediate checkpoint, and training is not finished

The same product update is explicit about the model’s status. “ERNIE 5.5 Flash Preview is an intermediate checkpoint from post-training,” Baidu writes. “It is already competitive with models of its size, but because training is not finished, some capabilities may still fall short.” Post-training is “ongoing,” it adds, and an official release is still to come. That warning should sit in front of the demos, not after them.

Those demos are extensive. Baidu shows long-horizon game-development tasks, Blender and FreeCAD modelling, research reports and video production, and says the model completed each piece end to end inside an agent harness, with external assets credited beside the work. The page also refers to scores “across major benchmarks.” This post does not repeat those numbers: the published appendix is not a usable, independently checkable table here, and a preview that may still “fall short” should not be introduced by a vendor scoreboard.

What is usable now is the API. Anyone with Qianfan access can call ernie-5.5-flash-preview at the documented 1 million-token window and the default rate limits above. The open weights, if and when they arrive under Apache 2.0, would be a later event, and should be recorded as an update to this post rather than as a fact of 11 November.

The 1 million-token window is the other hard fact on the product page and the Qianfan list. It is unusually long, and it is the main official specification besides the 280 billion / 14 billion split. Default rate limits of 60 requests per minute and 1 million tokens per minute are modest for a public preview endpoint; they are also the only capacity figures Baidu has published for this model. None of that tells a caller how the unfinished checkpoint will behave on a messy, real workload, which is why Baidu’s own “may still fall short” sentence is the one that matters.

A post-trained preview is not the same release as the open base

This is a new model endpoint, not a relabelling of last week’s files. TSN’s report on ERNIE-5.5-Flash-Base covered the earlier open-weight base. Flash Preview is the first post-trained model Baidu has put on top of that base, and it is being sold first as an API. The planned Apache 2.0 dump is a separate step, and it has no date attached.

It also sits in a wider run of Chinese open-model releases, including Baidu’s Unlimited-OCR 2, a document model from the same company, and DeepSeek’s decision to open-source DeepInfer, the engine behind its own services. Those are useful comparisons only in one narrow sense: other laboratories have already published weights or serving code, while Flash Preview’s weights are still a promise.

The honest reading of 11 November is therefore modest. A 280 billion-parameter mixture-of-experts model, with 14 billion parameters active and a 1 million-token context, is now callable on Qianfan. Baidu itself says the checkpoint is unfinished and that some capabilities may lag. Until the official release, and until the Apache 2.0 weights actually appear, that is the whole of the news.