Baidu Releases ERNIE 5.5 Flash Base as an Open Model
Baidu released ERNIE 5.5 Flash Base as an open model.
1 source 1h ago
Reporting from 1 source: GIGAZINE.
Why it matters
Baidu is betting that cost-controlled foundation models with long-context handling baked in at the base stage will be easier to extend than models patched after the fact, and it is shipping the base weights openly before the tuned version arrives.
Key Facts
- Baidu released ERNIE 5.5 Flash Base as an open model.
- ERNIE 5.5 Flash Base is a MoE model with 280 billion total parameters and 14 billion active parameters.
- ERNIE 5.5 Flash Base was pretrained on 33.5 trillion tokens and supports a 1 million token context window.
- Baidu plans to release a further-trained ERNIE 5.5 Flash in the near future.
- The release compares ERNIE 5.5 Flash Base against DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base.
It is a mixture-of-experts model with 280 billion total parameters and 14 billion active parameters, pretrained on 33.5 trillion tokens, with a 1 million token context window for inference and fine-tuning. Baidu says it is built so training and execution costs stay low, and that long-context understanding is built in at the foundation stage. A version with additional training, ERNIE 5.5 Flash, is planned for release soon.
Baidu published the weights on Hugging Face under the baidu/ERNIE-5.5-Flash-Base repository. The MoE design keeps 14 billion of its 280 billion parameters active per pass, which is the part that matters for execution cost.
Baidu frames the design around keeping additional training and inference cheap, and around adding abilities like long-context understanding at the foundation stage rather than later. Benchmark results in the release compare it against DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base, with ERNIE 5.5 Flash Base scoring above foundation models of comparable scale. A further-trained ERNIE 5.5 Flash is planned for release in the near future.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google