Baidu Releases ERNIE 5.5 Flash Base: A 280B MoE Base Model With 14B Active Parameters

Why it matters

Baidu has released ERNIE 5.5 Flash Base, a text-only sparse Mixture-of-Experts base model intended as a clean pre-training checkpoint for teams comparing fine-tuning, instruction tuning, or reinforcement learning methods. The weights stop after pre-training, with no mid- or post-training applied, so the model is a research starting point rather than a chat product. Baidu reports it beating DeepSeek-V4-Flash-Base on most base-model benchmarks in its unified evaluation pipeline, including knowledge tasks, while trailing on several others, notably reranking and long-context benchmarks where DeepSeek-V4-Flash-Base and a Tencent model lead. All figures are self-reported by Baidu, and competitor numbers measured in Baidu's pipeline differ slightly from what each vendor publishes on its own, so readers should treat the rankings as one lab's apples-to-apples comparison rather than settled ground truth. The release is notable for its architecture choices, including compressed attention layers with sparse KV caching, a multi-router expert design, and a final layer that doubles as a speculative-decoding draft model. Weights are available in FP8 and BF16 for SGLang and vLLM serving.