What Baidu released
Baidu has released ERNIE-5.5-Flash-Base, an open-source foundation model distributed deliberately without fine-tuning or post-training, according to The Decoder. The choice implies that those who use the model will decide its use and complete the subsequent adaptation stages; the source indicates that a post-trained version is expected later, without providing a timeline.
The model adopts a mixture-of-experts architecture with 280 billion total parameters, of which 14 billion are active for each input. Pre-training used 33.5 trillion tokens, with a training context of up to 128,000 tokens, while during fine-tuning and inference the model is designed to support a context of up to one million tokens.
Stated comparisons and availability of the weights
According to Baidu, in tests on reasoning, code and long-context tasks the model is competitive with other open-source foundation models of similar size, such as DeepSeek V4 Flash and MiMo V2.5. This is a company statement reported by the source, not a result verified by third parties: the cited results are not accompanied, in the available material, by detailed benchmark data.
The weights are available on Hugging Face in FP8 and BF16 formats. Publishing the weights in two formats is concrete information for those intending to evaluate or reuse the model, because it concerns the practical means of access, not just the announcement.
Why it matters and what remains to be verified
The release signals that Baidu's development of artificial intelligence models is continuing, as The Decoder writes, but with an approach that transfers much of the work to developers. Releasing a base model that has not been post-trained is a different choice from distributing a ready-to-use model: it shifts to the user decisions about data, objectives and adaptation procedures, and makes the availability of computational resources and fine-tuning expertise central.
The comparison with DeepSeek V4 Flash and MiMo V2.5 places ERNIE 5.5 Flash Base in a tier of open-source models of similar size, but the material does not allow establishing what the real performance differences are or under what conditions they were measured. Likewise, it does not indicate when the post-trained version will arrive, with what data or methods it will be produced, or what licenses accompany the weights published on Hugging Face.
At least three points therefore remain to be verified: the reproducibility of the results declared by Baidu on reasoning, code and long context; the actual functioning of the context up to one million tokens in the fine-tuning and inference stages; and the terms of use associated with the distribution of the weights in FP8 and BF16 formats.
