ERNIE 5.5 Flash Base: weights, deployment and license limits

Baidu announced ERNIE 5.5 Flash Base on November 5, giving model builders a downloadable starting point for further training. Its custom license excludes the European Union, United Kingdom and South Korea. Baidu’s announcement describes a text model taken directly from pretraining, with a version refined through further training still to come.

The immediate question is whether this checkpoint fits a particular development project. A team preparing specialized training data has a different decision to make from someone choosing an assistant for everyday work. Downloadable weights give that team more control over what happens next, while leaving it responsible for the training, evaluation and deployment work.

What the base checkpoint actually provides

Baidu lists 280 billion total parameters, with 14 billion activated per token, and 33.5 trillion pretraining tokens. The maximum pretraining context was 128K tokens; fine tuning and inference support up to 1M. FP8 and BF16 weights are available. These are specifications for a base model, not evidence of a finished conversational assistant.

For a researcher, starting before instruction tuning can be useful: competing training approaches can begin with the same checkpoint. For a product team, it creates additional work. The team still needs to decide how the model should respond, what examples teach that behavior and how to measure failures. A polished demo is a poor substitute for those decisions.

Likewise, a large context limit should be treated as a capacity to investigate. A useful evaluation would put relevant evidence at different positions in long inputs, include plausible distractions and check whether answers remain grounded in the supplied material. Passing a short prompt test says little about that workload.

Why parameter counts need context

The Hugging Face repository metadata reports 286,245,697,882 safetensors elements across its listed data types, rather than exactly 280 billion. The model card says its headline parameter counts exclude 0.36 billion additional ECHO parameters. That exclusion alone does not reconcile the figures, so the repository total should not be silently substituted for Baidu’s architectural count.

More importantly, activated parameters describe computation, not the complete memory requirement. An organization budgeting a deployment should work from the actual checkpoint, supported precision, serving configuration and expected request load. Treating a sparse model as if only its active parameters must fit into memory would give a misleading starting estimate.

Serving requires several GPUs

Baidu’s vLLM integration documents NVIDIA GPU support using vLLM 0.32.0 and its ERNIE plugin. Its published guidance lists roughly 280GB of FP8 weights or 560GB in BF16. For FP8, it lists two 288GB GPUs, four 192GB GPUs or eight 80GB GPUs as the smallest configurations that hold the weights. The documented setup with eight 80GB GPUs does not fit a 1M context.

Those figures are deployment guidance, not ByteForward measurements. We have not run this checkpoint or measured latency, throughput or memory use. They are nevertheless a useful warning against reading “14 billion active” as a consumer hardware recommendation.

The SGLang cookbook also separates low latency and high throughput recipes and tells operators to size context and concurrent requests for their workload. That is the right operational question: how many useful requests can a particular installation serve at an acceptable response time?

Before reserving a cluster, write down the expected input length, output length and concurrency. Then test a representative sample. A system optimized for one user working through a large document may need different settings from a service handling many short requests.

The custom license changes the adoption decision

The ERNIE 5.5 Flash Base Model License Agreement v1.0 limits its territory and extends that restriction to derivatives and outputs. Qualifying commercial use at or below US$50 million in related annual revenue has no model license charge, but requires a notice before first commercial use and prominent “Powered by ERNIE” attribution. Above that threshold, prior written authorization is required.

Revenue includes whole related products or services, not just the portion attributed to ERNIE, and aggregates entities under common control. The terms also restrict using the model or its outputs to improve other AI systems outside its derivative model allowance. Noncommercial use remains subject to the agreement. The file lists a November 5 release and November 6 acceptable use revision date. Users should confirm the applicable version.

Downloadable weights do not mean unrestricted use. Review the full terms before adoption. For a contrasting example, our Ripple-65B Preview guide explains a release under Apache 2.0.

Benchmarks are a starting point for evaluation

Baidu reports MMLU-Pro scores of 78.4 against 69.2 for DeepSeek V4 Flash Base in its shared evaluation setup. Its table also shows ERNIE trailing DeepSeek on ChineseSimpleQA. These are vendor reported comparisons of base checkpoints, not independent tests or results for the forthcoming assistant.

The repository history shows weight uploads before November 5, including November 2 and 3. November 5 is the dated announcement, not a verified date when the weights first became public.

For an eligible team with the infrastructure and a clear training objective, the release warrants a focused evaluation. Start with the license, then test the tasks that matter. A high score on someone else’s benchmark cannot decide whether the resulting model will serve your users well.

Featured image: AI generated editorial illustration by ByteForward, using the official Baidu and ERNIE logos. Conceptual artwork, not product hardware. No endorsement implied.