Baidu has published the weights of ERNIE-5.5-Flash-Base, a Mixture-of-Experts model with “280B total parameters and 14B activated per token”, on Hugging Face. Its license lets companies use it commercially without a model license charge while annual revenue of the related products or services is at or below US$50 million.
Above US$50 million, commercial use needs Baidu’s separate written authorization, the license says. It is dated November 5, 2026. By our arithmetic, 14B of 280B means 5.0% of the parameters are active for each token.
Business Pill · Distribution Advantage
A short explainer on distribution advantage. It teaches the general idea only and does not discuss this story.
The key insight: As we read Baidu’s own card and license, ERNIE-5.5-Flash-Base pairs a 280B-parameter base model, with 14B active per token (5.0% by our arithmetic), with a license that is free for commercial use up to US$50 million of related annual revenue and bars using its outputs to improve other AI models.
The model card says the model was “pre-trained from scratch on 33.5T tokens” with a training context of up to 128K, and that “fine-tuning and inference support a 1M-token context”. It is the checkpoint at the end of pre-training: the card calls it “the checkpoint at the end of pre-training, before any mid- or post-training”.
Baidu lists 43 layers, 512 routed experts per MoE layer (top-10) plus one shared expert, and a vocabulary of 201,216 tokens. FP8 and BF16 weights are both published.
MMLU-Pro (5-shot, exact match) for four base models as reported by Baidu on the ERNIE-5.5-Flash-Base model card: ERNIE 78.4, DeepSeek-V4-Flash-Base 69.2, Hy3-preview-Base 66.9, MiMo-V2.5-Base 65.7. These are Baidu’s own figures; the 9.2-point gap is FourWeekMBA arithmetic.
What Baidu’s Benchmarks Show
Baidu reports its own results against three similar-sized base models, all run through what it calls “the same pipeline, prompts and few-shot settings”. On MMLU-Pro it lists 78.4 for ERNIE against 69.2 for DeepSeek-V4-Flash-Base, 9.2 points higher by our arithmetic. On LiveCodeBench it lists 40.1 against 31.9.
ERNIE does not lead every row in Baidu’s table. It lists DeepSeek-V4-Flash-Base ahead on ChineseSimpleQA (75.3 vs 73.0) and InfiniteBench (48.0 vs 46.5), and Hy3-preview-Base ahead on FullStackBench (56.7 vs 54.6). These are Baidu’s numbers; we did not run the models.
The card lists DeepSeek-V4-Flash-Base at 284B total and 13B activated parameters, 4.6% active by our arithmetic, against ERNIE’s 5.0%.
What the License Requires
Non-commercial use, including research and personal experimentation, is free with no notice. Commercial users at or below the US$50 million revenue line must send Baidu a one-time use notice and display “Powered by ERNIE” in at least one readily accessible place, such as the product’s documentation.
The license also bars using its outputs to train other systems. Section 5.5 reads: “You may not use the Licensed Materials, any Derivative Model, or any of their Outputs or results to improve any other artificial intelligence model or machine-learning system other than the Model or its Derivative Models”.
As we read the card, Baidu released the checkpoint at the end of pre-training, before any mid- or post-training, so builders pick their own fine-tuning path.
As we read Baidu’s own table, ERNIE leads most rows but not all: Baidu lists DeepSeek-V4-Flash-Base ahead on ChineseSimpleQA and InfiniteBench, and Hy3-preview-Base ahead on FullStackBench.
As we read the license, the revenue line, the “Powered by ERNIE” notice and the Section 5.5 limit on training other models are the terms that decide commercial use.
“by releasing the weights at the end of pre-training, we let developers build on the model’s general capabilities and choose for themselves where to take it next”
Three Implications
FOR BUILDERS Non-commercial use is free; commercial use at or below US$50 million of related annual revenueneeds a one-time notice and a Powered by ERNIE display, per the license.
FOR LARGER COMPANIES Above US$50 million, commercial use needs Baidu’s separate prior written authorization, per the license.
FOR READERS The benchmark figures are Baidu’s own; we did not test the model.
The analysis says: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, Baidu opens the weights while keeping the commercial right above US$50 million and the right to train other models on its outputs. The analysis does not discuss Baidu or ERNIE.
What Is Not Established
The benchmark figures are Baidu’s own and we did not test the model. The card does not say what hardware the model was trained on or what training cost. We did not read Baidu’s technical blog, and we did not contact Baidu.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Baidu’s ERNIE-5.5-Flash-Base is a 280B-parameter model with 14B active per token, trained on 33.5T tokens, and its weights are free to use commercially up to US$50 million in related annual revenue, with a “Powered by ERNIE” notice. Baidu’s own table puts it 9.2 points ahead of DeepSeek-V4-Flash-Base on MMLU-Pro, by our arithmetic.
A note on sourcing. Model details, benchmark figures and license terms are from Baidu’s own model card and license on Hugging Face; the benchmarks are Baidu’s and we did not test the model. The percentages and the point gap are our arithmetic. We did not contact Baidu.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
