Baidu open-sources ERNIE 5.5 Flash Base, a 280B model that comes with no fine-tuning at all

Baidu has released a new open model, and left most of the finishing work to you. ERNIE-5.5-Flash-Base ships as a raw foundation model with no fine-tuning or post-training, Matthias Bastian reports for The Decoder. Developers decide where to take it; Baidu says a post-trained version will follow.

The specs

• 280 billion parameters in a mixture-of-experts design, with 14 billion used for any given input.
• 33.5 trillion pretraining tokens. Training used contexts as long as 128,000 tokens, and the model is built to handle up to a million during fine-tuning and inference.
• Weights on Hugging Face in FP8 and BF16.

Baidu says it holds its own against similar open base models, including DeepSeek V4 Flash and MiMo V2.5, on reasoning, code and long-context benchmarks. Those are Baidu's numbers from its own evaluation pipeline, so wait for independent runs.

Who this is for

A base model isn't a chatbot. Prompt it like one and you'll get autocomplete, not answers. It's for teams that want to do their own instruction tuning or domain training, and who would rather start from a clean pretrained checkpoint than undo somebody else's alignment choices. Fourteen billion active parameters also keeps inference cost closer to a mid-size dense model, though you still need room for all 280 billion weights.

Details are thin beyond that: the report doesn't cover the license terms or when the post-trained release lands, so check the model card before you plan around it.

Scoreboard update: AI +1. Another 280B open model lands, ready for anyone to fine-tune.

Stay undead.