- Baidu released ERNIE-5.5-Flash-Base, an open foundation model that ships without fine-tuning or post-training.
- The model uses a mixture-of-experts architecture with 280 billion total parameters, 14 billion active per input.
- It was pre-trained on 33.5 trillion tokens and is designed to support context windows up to one million tokens after fine-tuning.
- Baidu says the model competes with similarly sized open models like DeepSeek V4 Flash and MiMo V2.5 on its own benchmarks.
What Happened
Baidu released ERNIE-5.5-Flash-Base, an open-weight foundation model that ships deliberately without fine-tuning, leaving developers to decide how to adapt it, The Decoder reported Nov 5. Baidu says a post-trained version is expected to follow. The weights are available on Hugging Face in both FP8 and BF16 formats.
Why It Matters
The release continues Baidu’s run in open-weight models after its ERNIE line was already judged to be among the models that had surpassed DeepSeek by UBS’s estimate. It also adds to a broader wave of Chinese open-weight releases, following moves like DeepSeek open-sourcing its DeepInfer inference engine, as Chinese labs compete partly on how openly they release their work.
Technical Details
ERNIE 5.5 Flash Base uses a mixture-of-experts design with 280 billion total parameters and 14 billion active per input, pre-trained on 33.5 trillion tokens with a 128,000-token training context that Baidu says can extend to one million tokens during fine-tuning and inference. Baidu says the model competes with similarly sized open base models — DeepSeek V4 Flash, Hy3-preview-Base, and MiMo V2.5 — across reasoning, code, and long-context benchmarks, using the same evaluation pipeline for all four.
Who’s Affected
Developers who want a base model to customize rather than a ready-to-use chat assistant are the primary audience, since ERNIE 5.5 Flash Base has no instruction-tuning out of the box. Other open-model labs, including DeepSeek and Xiaomi’s MiMo team, are the direct benchmark comparisons Baidu chose to highlight.
What’s Next
Baidu has not given a release date for the post-trained, instruction-following version of ERNIE 5.5 that it says will follow this base model.
Source: The Decoder, Nov 5, 2026.
