Baidu releases ERNIE as a training foundation

A checkpoint for further work

Baidu’s model card identifies ERNIE 5.5 Flash Base as the checkpoint after pre-training, before subsequent training stages. It is intended as a starting point for domain adaptation, instruction tuning and reinforcement-learning research. The text-only model’s weights are offered in FP8 and BF16 formats. Developers receive a foundation for shaping behaviour, rather than a claim that their application’s training is already complete.

Running the released model

Baidu’s serving guide distinguishes the released Base model, which continues text, from a forthcoming chat version with reasoning and tool calling. The plugin runs with unmodified vLLM 0.32.0 on NVIDIA GPUs only. Base has not received chat training. Its prompts go to the completions endpoint. The sample server commands bind to all interfaces without authentication; the guide advises an API key or a local-only binding unless the network is trusted. This is a documented deployment recommendation to consider, not evidence that a downloaded model is automatically a protected service.

Why the distinction matters

The report frames the release as leaving substantial adaptation work to developers. That is the report’s interpretation of the offering: the useful starting point still leaves teams to decide what they want to build and how to train it.