Baidu promises looser terms for its next free AI model, a week after barring Europe from the base version

Baidu says it will release the downloadable files for its newest AI model, ERNIE 5.5 Flash Preview, under Apache 2.0. That is a standard license that lets any company use, change and sell products built on a model, anywhere, without asking permission or paying fees.

For now that is a promise, not a license. If Baidu follows through, it will undo the terms it set a week ago. The base version of the same model, released November 5, cannot be used in the European Union, Britain or South Korea. Firms whose related products earned more than $50 million in the past year also need Baidu's written approval.

For buyers, that gap is large. Under the base license, a bank in Frankfurt or a manufacturer in Seoul cannot run the model. Apache 2.0 has none of those clauses. It also lacks the base license's demand for a visible "Powered by ERNIE" label and its ban on using the model's answers to improve other AI systems.

The catch is that nothing has shipped. As of November 12, the preview was available only as a paid service on Baidu's Qianfan cloud and as a free listing on OpenRouter, a marketplace that sends requests to many AI models. Reports through that day said no downloadable files had been posted.

What Baidu released, and what it promised

ERNIE 5.5 Flash Preview, announced November 11, is the first version of the model given the extra training that turns raw material into a usable assistant. The base model released a week earlier was unfinished, meant for companies to customize. Baidu calls the preview "an intermediate checkpoint," says training is still under way and promises improvements in an official release.

The model is large but built to be cheap to run. It has 280 billion parameters, the internal settings a model learns in training, but uses only 14 billion of them for each word it writes. It can read about a million tokens at once, roughly 750,000 words.

Sources disagree on when, and on what terms

The timing is less clear than the license. Baidu's English-language page says the files will be released under Apache 2.0 "soon," with no date. ModelFit, a site that recommends models to run on personal hardware, reports that Baidu's Chinese-language page promises them before November 25. Several outlets repeat that date. TechNode, which covers Chinese tech, reported no date.

Not every outlet reads the terms the same way. AlphaSignal, an AI news site, warned that Apache 2.0 assumptions do not apply, without making clear whether it meant the preview or only the base model. Until Baidu posts the actual license text with the files, the promise is all there is.

Prices are also unsettled. Baidu's announcement lists none, and third-party figures for its Qianfan service differ. OrcaRouter, a company that resells access to AI models, lists $0.28 per million input tokens. Superpower Daily, an AI newsletter, reports a lower rate in Chinese yuan.

How good is it? Mostly Baidu's word

Baidu published 20 test results, run on its own setups, some graded by rival AI models it chose as judges, according to OrcaRouter. None has been reproduced by an outside group. Artificial Analysis, an independent testing firm, had not listed the model as of November 11.

The results are mixed even on Baidu's numbers. On AutomationBench, a test of office-style tasks, Baidu reports 61.9%, ahead of Anthropic's Claude Opus 5.5 at 58%. On Terminal-Bench 4.0, which measures work done at a computer's command line, it reports 46.5%. That trails Claude Opus 5.5 at 63.6% and OpenAI's GPT-6.1 Sol at 56.1%.

The one outside signal is provisional. According to OrcaRouter, Arena, a popular model-ranking site, scored the preview 1668 on its web-building leaderboard, using automated scoring rather than human votes. Arena said the model was "trending towards" becoming the second-best open model.

OrcaRouter, which sells access to many of the rival models, disputed that reading. It noted the entry is tagged proprietary, would rank around 17th overall, and is statistically tied with several rivals, including models from DeepSeek and Moonshot.

The model also handles only text. In one demo, Baidu used a tool powered by Anthropic's Claude Opus 5.5 to handle images.

Why the license matters more than the scores

Baidu gave no reason for the base model's regional limits, Trending Topics reports. Its previous free release, ERNIE 4.5 in June 2025, used Apache 2.0. The preview promise is a return to that approach.

Rivals already offer easier terms. DeepSeek released its V4 Flash model under the MIT license, which is as permissive as Apache 2.0, and Alibaba's Qwen models use Apache 2.0, according to Trending Topics. A license that shuts out the EU and Britain gives those customers a reason to pick a rival.

Alibaba is also launching services first and leaving the terms for later. It opened a preview of its Qwen4-Max model the same day with no downloadable files, no license and no published test results, according to Fello AI, which sells an app bundling several AI chatbots.

Three things will show whether Baidu's promise is real. First, whether the files arrive by November 25. Second, whether the license text is plain Apache 2.0, with no added clause on regions or revenue. Third, whether Baidu moves the base model onto the same terms.

Until then, companies in the EU, Britain and South Korea still have no downloadable ERNIE 5.5 model they can legally use. A promised license is not a license.