Baidu releases open-weight ERNIE 5.5 Flash Base model
Baidu published weights for ERNIE 5.5 Flash Base, a 280B-parameter model with a restrictive license.
Objective Facts
Baidu open-sourced the weights of ERNIE 5.5 Flash Base on Nov. 5, 2026, posting the model on Hugging Face, according to QbitAI and PingWest. Xinhua reported the model uses a mixture-of-experts design with 280 billion total parameters and 14 billion active per token, was pretrained on 33.5 trillion tokens, and supports up to 1 million tokens of context in fine-tuning and inference. Baidu says the base model beats DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base among models of similar size, though Unite.AI's table of Baidu's own figures shows rivals ahead on some tests. The license bars use in the EU, UK and South Korea and requires separate authorization for larger commercial users, Unite.AI and XenoSpectrum reported. Chinese coverage largely repeated Baidu's performance claims, while Western and Japanese tech sites emphasized the license terms and the GPU memory needed to run the model.
Deep Dive
ERNIE 5.5 Flash Base is a pretraining checkpoint, released before instruction tuning or reinforcement learning. Unite.AI reported that Baidu pitched it as a shared starting point for domain adaptation and training research, and as a fixed baseline for controlled comparisons. The model has 43 layers, and 37 of them compress attention so that each block of 128 tokens takes one cache entry, which Baidu says keeps long-context cost in check, according to Unite.AI and XenoSpectrum. Baidu has released open models before; RuntimeWire listed the ERNIE 4.5 300B-A47B base model as the earlier comparison and said the new license is tighter than the ERNIE 4.5 terms. Baidu's performance claims are partly supported by its own numbers. Unite.AI reported that ERNIE scored 93.3 on MMLU and 82.3 on HumanEval+, ahead of DeepSeek V4 Flash Base, but Hy3 preview led on FullStackBench and DeepSeek led on InfiniteBench. XenoSpectrum added that ERNIE trailed DeepSeek on ChineseSimpleQA (73.0 vs. 75.3) and HELMET-Rerank (56.2 vs. 60.0). None of the outlets reviewed reported independent testing. XenoSpectrum's Y Kobayashi also cautioned that the 5% active-parameter figure describes compute per token, not memory: the weights take about 280GB in FP8 and about 560GB in BF16, and Baidu's 1-million-token vLLM example uses four GPUs with 184GiB each. The license will shape who can use the model. Under the ERNIE 5.5 Flash Base Model License Agreement v1.0, as described by Unite.AI, non-commercial use is free, companies with up to $50 million in related revenue need a one-time notice and "Powered by ERNIE" attribution, and larger firms need Baidu's prior written permission. The license also bars using the model or its outputs to improve unrelated AI systems, which Unite.AI said matters for distillation and synthetic-data work. The next items to watch are the post-trained ERNIE 5.5 Flash release, for which XenoSpectrum said Baidu has set no date, and whether outside researchers reproduce Baidu's benchmark results.
Regional Perspective
In China, Xinhua reported on Nov. 5 that Baidu had officially open-sourced ERNIE 5.5 Flash Base, listing its 280 billion parameters, 14 billion active parameters and 1-million-token context, and describing efficiency as the model's central goal. QbitAI and PingWest published nearly the same text the same day under a headline saying the model set new same-size records across several benchmarks; QbitAI said the content came from Baidu and was reprinted with permission. Both reports described Baidu's technical changes, including a new sparse attention design, a multi-router expert system and a training objective called ECHO. The Chinese reports did not mention the license's exclusion of the EU, UK and South Korea or the $50 million commercial threshold, which Unite.AI and RuntimeWire treated as central. They also presented the comparison with DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base as a win, while Unite.AI's table showed rivals leading on FullStackBench and InfiniteBench. In Japan, XenoSpectrum's Y Kobayashi wrote that the model is a base for fine-tuning rather than a ready chatbot, and that companies should check revenue calculations and regional rules before use. Xinhua, QbitAI and PingWest did not report independent testing of Baidu's comparisons. For Japanese businesses, XenoSpectrum framed the questions as practical ones: whether they fall under the $50 million line, and whether they have the GPU memory to run a model whose full weights must be loaded even though only about 5% are active per token.