Baidu has released ERNIE-5.5-Flash-Base, an open-source foundation model containing 280 billion parameters. This specific release ships without fine-tuning or post-training, leaving developers to determine their own use cases while a post-trained version is planned for later. The architecture employs a mixture-of-experts design where 14 billion parameters remain active per input. Pre-training consumed 33.5 trillion tokens and supports a context window of 128,000 tokens, though the system is designed to handle up to one million tokens during fine-tuning and inference. Weights are currently available on Hugging Face in FP8 and BF16 formats.
The model targets performance in reasoning, code generation, and long-context handling, according to Baidu. Benchmarks suggest it competes with similarly sized open foundation models like DeepSeek V4 Flash or MiMo V2.5. This release adds another large-scale option to the open-source ecosystem without altering existing customer products or API behaviours. The strategy allows third parties to adapt the base model for specific requirements before official fine-tuned versions arrive.
- 280 billion total parameters with 14 billion active per input
- Pre-trained on 33.5 trillion tokens
- Context support up to one million tokens during fine-tuning
