Baidu releases ERNIE 5.5 Flash Base, a 280B open foundation model
Baidu has released ERNIE-5.5-Flash-Base, an open foundation model that ships deliberately without fine-tuning or post-training. The Decoder's framing is that Baidu's AI development is still alive, but most of the work is left to the people who download the model: developers are expected to decide for themselves where to take it. A post-trained version is supposed to follow.
ERNIE 5.5 uses a mixture-of-experts architecture with 280 billion parameters in total, 14 billion of which are active per input. It was pre-trained on 33.5 trillion tokens. The training context is up to 128,000 tokens, though the model is designed to support a context of up to one million tokens during fine-tuning and inference.
On benchmarks for reasoning, code and long-context tasks, the model competes with similarly sized open foundation models such as DeepSeek V4 Flash or MiMo V2.5, according to Baidu. The weights are available on Hugging Face in FP8 and BF16.
Key facts
- Baidu released ERNIE-5.5-Flash-Base, an open foundation model shipped without fine-tuning or post-training; a post-trained version is supposed to follow.
- It is a mixture-of-experts model with 280 billion parameters in total, 14 billion of them active per input.
- Pre-training used 33.5 trillion tokens with a training context of up to 128,000 tokens; the design target is up to one million tokens during fine-tuning and inference.
- Baidu says the model competes with similarly sized open foundation models like DeepSeek V4 Flash or MiMo V2.5 on reasoning, code and long-context benchmarks.
- Weights are on Hugging Face in FP8 and BF16.
Why it matters
A 280-billion-parameter mixture-of-experts model with only 14 billion parameters active per input is a large open release from a Chinese company. Baidu is positioning it against similarly sized open foundation models such as DeepSeek V4 Flash and MiMo V2.5. The release is a base model on purpose: it comes without fine-tuning or post-training, so the shaping of its behaviour is left to developers.
Who it affects
Developers and teams who want to build on an open base model and are ready to fine-tune it themselves. The Decoder says developers are expected to decide for themselves where to take the model. Those who want a ready-to-use assistant will need to wait for the post-trained version Baidu is supposed to release.
How to use it
The weights are available on Hugging Face in FP8 and BF16. The model is designed to support a context of up to one million tokens during fine-tuning and inference, although its training context is up to 128,000 tokens. The source does not state the license, pricing, API availability or hardware requirements.
How solid is it
The architecture, parameter counts, token count, context lengths and weight formats come from the report on the release itself. The competitiveness claim is Baidu's own: the model competes with DeepSeek V4 Flash or MiMo V2.5 "according to Baidu". No benchmark scores or numeric comparisons are given, and the source does not say which benchmarks were used.
Risks and caveats
No release date is given, and there is no date or timescale for the promised post-trained version. The license of the model is not stated. Because this is a base model without post-training, the benchmark claims are Baidu's and cannot be checked from this report, and the one-million-token context is a design target rather than the training context.