Baidu releases raw 280B AI model weights with limits on commercial use

Developers can now start from Baidu’s large-scale pretraining work and choose their own next training steps. On November 5, 2026, Baidu released ERNIE 5.5 Flash Base, a downloadable 280-billion-parameter language model. It arrives before mid-training or post-training, giving teams a shared starting point rather than Baidu’s finished assistant behavior.

A starting point, not a finished assistant

Baidu positions the checkpoint for domain adaptation, instruction tuning and reinforcement-learning research. Those are different ways to add specialist knowledge or shape how a model responds. Releasing this earlier stage lets developers choose their own training goals instead of starting with a model that has already undergone Baidu’s later training stages.

There is also a research purpose: teams can compare later training methods from the same checkpoint and measure what each adds. Baidu says the release is intended to spare developers the cost of large-scale pretraining, leaving them to spend resources on the tasks and training methods they care about.

The weights are available on Hugging Face in FP8 and BF16 numerical formats. Baidu documents serving through SGLang and vLLM using a completions interface and points developers to an ERNIE 5.5 example for fine-tuning. The release therefore includes routes for both running the checkpoint and continuing its training.

280 billionTotal parameters

The model’s full parameter pool, according to Baidu.

14 billionActive per token

Its sparse mixture-of-experts design activates only part of the model for each token.

Where Baidu cuts the work

The mixture-of-experts design routes each token—a unit of text—through selected specialist networks rather than the entire parameter pool. Each expert layer has 512 routed experts, selects 10 and adds one shared expert. Baidu uses several routing networks to make those selections, saying this improves expert choice without increasing the number activated.

Long inputs get a separate efficiency treatment. Most attention layers compress every 128 tokens into one cached entry while retaining a short window of recent, uncompressed text. One layer in seven retains token-level access across the full context. In those layers, a lightweight indexer selects positions for the main attention calculation.

Baidu says the model supports a one-million-token context for fine-tuning and inference—the amount of material it can work with at once. That differs from its pretraining context, which grew to 128K tokens. The checkpoint was pretrained from scratch on 33.5 trillion tokens.

Another component, ECHO, reuses the model’s final layer to predict future tokens instead of attaching a separate prediction module. Baidu says this trains the same parameters used for next-token generation. The layer also drafts candidate tokens for speculative decoding, a process in which the model checks proposed text before accepting it.

The training data extends beyond standalone code files. Baidu included commits, issues, pull requests, code reviews, generated tests and fixes checked through continuous integration. Its long-context data combines natural documents, related texts mixed with distractors, and synthetic retrieval and question-answering tasks. These are deliberate targets for later repository-scale coding and long-document work.

For evaluation, Baidu compared the released checkpoint with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base. It says all four used the same evaluation harness, prompts and few-shot examples, plus the same model-based judge for free-form answers. Baidu highlights reasoning, code and exam results; these are company-run base-model comparisons, not a ranking of finished assistants.

Downloadable weights, conditional rights

Access to the files comes under the custom ERNIE 5.5 Flash Base Model License Agreement v1.0. Its geographic grant excludes the European Union, the United Kingdom and South Korea. Commercial rights also depend on revenue from related products and services, with different requirements on either side of a $50 million annual threshold.

• At or below $50 million: commercial use carries no model license charge, subject to attribution and a one-time notice to Baidu.
• Above $50 million: commercial use requires separate written authorization from Baidu before that use.

The revenue definition is consequential. It uses a rolling 12-month period and counts a related product’s entire revenue even when that product also uses other models or technologies. Entities under common control are aggregated. The license also includes adapters, fine-tuned models and models trained through distillation or synthetic-data knowledge transfer in its definition of derivatives.