Baidu open-sources ERNIE 5.5 Flash Base: a 280B-parameter MoE foundation model with no post-training, FP8 and BF16 weights on Hugging Face
On 5 November Baidu open-sourced ERNIE 5.5 Flash Base (ERNIE-5.5-Flash-Base), a Mixture-of-Experts model with 280B total and 14B active parameters per token. It is the end-of-pre-training base checkpoint, with no fine-tuning or post-training, and the weights are on Hugging Face in FP8 and BF16.
According to the model card, the model was pre-trained on 33.5 trillion tokens with a training context of up to 128K and supports up to 1 million tokens for fine-tuning and inference. Baidu says it releases the weights at the end of pre-training as a common starting point for domain adaptation, instruction tuning and reinforcement-learning research; The Decoder notes a post-trained version is expected later, so the current model is not a ready-made chat or agent model.
Under one evaluation pipeline, Baidu compares it with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base and claims same-size state-of-the-art results across knowledge, reasoning, math, code, multilingual and long-context benchmarks. The comparisons were run by Baidu itself, and base-model scores do not equal final product performance.
Baidu’s article, republished by QbitAI, says the generation was designed around efficiency, with changes to attention (Latent-Gated Sparse Attention), expert routing (Multi-Router Latent MoE) and the training objective (ECHO) to cut long-context memory and compute cost. On licensing, the model card says it is provided under the ERNIE 5.5 Flash Base Model License Agreement v1.0; commercial terms require reading that agreement.
Sources
- baidu/ERNIE-5.5-Flash-Base — Hugging Face（百度模型卡）
- 百度开源文心5.5 Flash Base，多基准测试刷新同尺寸模型 SOTA — 量子位（百度提供、授权转载）
- Baidu releases ERNIE 5.5 Flash Base with 280 billion parameters as an open-source foundation model — The Decoder