A downloadable model whose license excludes the European Union

Baidu’s ERNIE 5.5 Flash Base model is released for adaptation to other uses. Its license excludes the European Union, including France.

The parameters of ERNIE 5.5 Flash Base are available for download, but their release does not amount to authorization for use everywhere. The license attached to the files excludes the European Union, the United Kingdom, and South Korea. This territorial restriction applies to the model, its derivatives, and their outputs.

The release provides a starting point for teams wishing to continue training an already pretrained model. The files published on Hugging Face correspond to a “Base,” text-only version with 280 billion parameters in total, 14 billion of which are activated per token. The post-trained version of ERNIE 5.5 Flash is announced for later. An opening with terms of use attached

The release falls under open weights with a restrictive license. In the authorized territories, commercial use remains royalty-free up to 50 million dollars in annual revenue related to the products or services concerned, subject to notification and the mention “Powered by ERNIE.” Above that threshold, prior written authorization is required.

The contract also prohibits using the model or its outputs to improve other AIs, except the model itself and its derivatives in compliance with the contract. A base to adapt, before the assistant

The interest of this release lies in the stage at which it appears. The model comes out of pretraining, without intermediate training or post-training. Teams can work on adapting it to a domain or teaching it instruction following. It is not yet the full conversational assistant announced by Baidu.

Pretraining covered 33,500 billion tokens, with sequences reaching 128,000 tokens. The published configuration allows up to one million tokens for inference and subsequent training. These two values describe different stages: the maximum announced capacity does not correspond to the length used during pretraining.

The weights are offered in FP8 and BF16, two numerical precision formats that affect the memory required. The vLLM deployment guide specifies, for example, eight 80 GB H100 GPUs for its FP8 configuration. The 14 billion active parameters describe the portion used in the computation, not the size of all the weights to host. Keeping context without keeping everything at the same level of detail

Memory management occupies a central place in the technical post. Of the model’s 43 layers, 37 retain a highly compressed history, with one memorized entry for 128 tokens. The other six keep more detailed access to older tokens and select the positions useful for computation. The available memory is thus concentrated in a small part of the layers to retrieve distant information.

Expert selection is also redesigned. Several routing mechanisms jointly evaluate the sub-networks to be used, rather than a single decision-maker. According to Baidu’s tests, this organization improves expert selection without increasing their active number or the amount of computation devoted to them. The claimed gain concerns selection, without activating all 280 billion parameters.

A third modification, called ECHO, reuses the last layer to learn to anticipate additional tokens. This layer is also used to propose a draft of the rest of the text in speculative decoding, before verification. The method aims to speed up generation without resorting to a separate draft model. These mechanisms describe the architecture’s efficiency choices; by themselves, they do not provide an operating cost for a given application. Scores compared with those of other base models

The published evaluations pit this version against DeepSeek V4 Flash Base, Hy3 preview Base, and MiMo V2.5 Base, of similar sizes. The four models are tested by Baidu in the same setup, with the same instructions and examples for each benchmark. The comparison concerns base versions, before the stages that shape an assistant’s behavior.

On HumanEval+, dedicated to code, the announced score reaches 82.3, compared with 69.5 for DeepSeek V4 Flash Base. The table does not, however, give ERNIE the advantage on all benchmarks: on FullStackBench, it gets 54.6, behind Hy3 preview Base’s 56.7. The results come from the team publishing the model and do not constitute independent validation of general superiority.

The code corpus described by the researchers also includes project evolution: successive modifications, issue reports, code reviews, and fixes verified by automated tests. Pretraining therefore incorporates traces of work on a software repository, in addition to the code files themselves.