Baidu has released an open foundation model with 280 billion total parameters—featuring 14 billion active parameters—while deliberately omitting any pre-training fine-tuning or post-training steps, leaving enterprise adopters to absorb the full cost of adaptation.
The company published the system under the name ERNIE-5.5-Flash-Base on Hugging Face in FP8 and BF16 formats. While a post-trained version of ERNIE 5.5 is scheduled for a later release, engineering teams currently face the operational burden of establishing their own post-training pipelines for the raw weights. Post-training a model of this scale demands substantial compute infrastructure, including clusters of high-performance accelerators such as H100 or equivalent GPUs to handle the memory bandwidth and gradient computations required for alignment and fine-tuning. While exact dollar figures are absent from Baidu's disclosure, enterprise adopters must budget for thousands of GPU-hours and significant electricity and cluster rental expenses to bridge the gap left by the raw weights.
According to Baidu's technical disclosures, the model was pre-trained on 33.5 trillion tokens with an initial training context of up to 128,000 tokens. The underlying architecture supports a context window of up to one million tokens during both fine-tuning and inference stages, demanding significant memory bandwidth and compute clusters from deploying organizations.
Baidu positions the model to compete with open foundation models such as DeepSeek V4 Flash and MiMo V2.5 across standard benchmarks for reasoning, code generation, and long-context processing tasks.
This distribution strategy allows Baidu to optimize its own compute expenditures, offloading the final stages of refinement onto the broader community and enterprise clients.
Side by side
Who actually benefits from releasing raw foundation weights stripped of final alignment? Baidu trims its own compute expenses, leaving clients to cover the server bills for NVIDIA H100 clusters and custom post-training pipelines. Instead of a turnkey deployment, enterprise adopters get a raw base trained across 33.5 trillion tokens.