Baidu ERNIE 5.5 Flash Base: Weights, Hardware and License Limits

Baidu announced ERNIE 5.5 Flash Base on November 5, publishing the technical explanation for a downloadable checkpoint intended for further training. The official release describes a text model taken directly from the end of pre-training, before mid-training or post-training. A post-trained ERNIE 5.5 Flash is planned, without a release date in the announcement.

That makes this release relevant to teams developing specialized models rather than buyers looking for a finished chat assistant. Before allocating compute, there are two practical checks: whether the infrastructure can run it and whether the intended deployment fits its custom license.

What Baidu released

The model card lists FP8 and BF16 weights on Hugging Face and ModelScope. Baidu describes the mixture-of-experts model as having 280 billion total parameters and 14 billion activated per token, trained on 33.5 trillion tokens. Maximum pre-training context was 128K; fine-tuning and inference support up to 1,048,576 tokens.

Baidu's evaluations compare base checkpoints under a shared harness and prompting setup. Its reported HumanEval+ score is 82.3 versus 69.5 for DeepSeek-V4-Flash-Base, while InfiniteBench is 46.5 versus 48.0. These vendor-run results show both gains and exceptions. They do not establish how a later instruction-tuned version will perform as a coding agent.

For a training team, the useful experiment starts with the target task and a held-out evaluation set. Measure the base checkpoint before adaptation, then use the same tests to judge what additional training changes. A stronger aggregate score can hide regressions in the specific documents, languages or workflows a product needs.

Sparse computation still needs substantial memory

The architecture combines heavily compressed attention with a smaller number of layers retaining token-level access. Baidu's technical explanation says most layers compress each group of 128 tokens into one cached entry, while retaining a short recent-token window. This is part of its effort to reduce long-context memory and computation.

The active-parameter count should not be mistaken for the amount of weight memory required. Baidu's vLLM plugin documentation estimates roughly 280 GB for FP8 weights and 560 GB for BF16. It runs with vLLM 0.32.0 on NVIDIA GPUs. The plugin is Apache-2.0 licensed; that does not change the model's separate terms.

The serving guide lists eight 80 GB GPUs for FP8, but says that configuration does not fit the full million-token context. Long-context prefill needs additional temporary memory, and concurrency must be sized around the KV cache. The released base model uses the completions endpoint; the chat, reasoning and tool-calling model remains marked coming soon.

Use the documented configurations as starting points for capacity planning, then measure the actual workload. Weight size, context length and simultaneous requests all affect whether a serving configuration is viable. No deployment or throughput test was performed for this article.

The license limits where and how it can be used

The ERNIE 5.5 Flash Base license excludes the European Union, United Kingdom and South Korea from its permitted territory. Its restrictions extend to derivative models and outputs, so an internationally available product needs a careful review.

For qualifying commercial use with annual related-product or service revenue at or below US$50 million, the agreement requires a one-time notice before use and prominent “Powered by ERNIE” attribution, without a model license charge. Above that threshold, separate written authorization is required in advance. The calculation includes the full revenue of related products and aggregates entities under common control.

Non-commercial use has no model license charge, but remains subject to the agreement. The derivative-model definition includes certain distillation and synthetic-data training uses. Teams should read the full terms with counsel rather than assume downloadable weights carry permissive open-source rights.

The linked revision was checked on November 5. Its acceptable-use exhibit says November 6, a date inconsistency to clarify with Baidu before relying on it.

This release gives researchers and model builders a substantial new starting point. Its value will depend on the quality gained through further training, the cost of serving the resulting model and whether its distribution terms fit the product. Those decisions should come before a large training run.

I'm Paras. I write about AI models and the engineering choices behind their adoption.

I help technical founders turn complex releases into clear technical writing, launch stories and founder-led distribution.

Book a 30-minute call.