• Sources: primary, discussion
  • Summary: ERNIE-5.5-Flash-Base is released on Hugging Face in FP8 and BF16 with no mid- or post-training, a 128K training context and 1M-token support for fine-tuning and inference. Six of 43 layers use full-context sparse MLA attention while the rest compress history 128 times. It ships under a custom ERNIE model license and is served by SGLang and vLLM.
  • Why it matters: Post-training teams get a large open base checkpoint for controlled SFT and RL comparisons without paying for pre-training.