Baidu releases ERNIE 5.5 Flash Base weights: 280B-A14B MoE base model with 1M context
Released Nov 5, the pre-training-only checkpoint (33.5T tokens, no instruction tuning) ships in FP8 and BF16 on Hugging Face and ModelScope, serves via SGLang or vLLM with up to 1M-token context, and self-reports 82.3 on HumanEval+ versus 69.5 for DeepSeek-V4-Flash-Base. It's under Baidu's own model license, not an OSI license.