Baidu releases ERNIE 5.5 Flash: 280 billion parameters, but only 14 billion active

🤖 Baidu releases ERNIE 5.5 Flash: 280 billion parameters, 14 billion active

Baidu Research has released the open MoE model ERNIE-5.5-Flash-Base, trained on 33.5 trillion tokens: hybrid Latent-Gated Sparse Attention, MoE with 512 experts, and the ECHO mechanism — an embedded draft model for speculative decoding. Context: 128K during pre-training and up to 1M tokens, with weights in FP8 and BF16 already available on Hugging Face.

🌍 For business, this is a low-cost entry point into the open 280B MoE: fine-tuning for specific tasks and agentic scenarios without a proprietary API, while the optimized KV-cache reduces inference costs for long contexts.

👤 Weights can be run via SGLang or vLLM, and ECHO accelerates generation without a separate draft model. According to vendor benchmarks: MMLU 93.3, GPQA 60.2, HumanEval+ 82.3.

Source 1: https://ernie.baidu.com/zh/release/ernie-5-5-flash-base/ Source 2: https://huggingface.co/baidu/ERNIE-5.5-Flash-Base