Companies weighing open Chinese models get a second 280B Ernie checkpoint in a week, with Apache 2.0 terms promised, looser than the base model's custom license.
Baidu released ERNIE 5.5 Flash Preview on November 11, the first post-trained version of the Ernie base model it opened up six days earlier, and said the weights will be made public before November 25. The model has 280 billion parameters in total, of which 14 billion are active for each token, and a context window of one million tokens, according to Baidu’s release page. The API is already live on Qianfan, Baidu’s model platform, and the English announcement adds a listing on OpenRouter, including a free variant.
Baidu labels the release a work in progress. Its Chinese page calls the Preview "an intermediate version in the post-training process" (后训练过程中的一个中间版本) that, while competitive with models of its size, "may still have some capability deficits due to insufficient training". IT Home reported that Baidu will watch real-world performance and user feedback and improve the model in a formal release, for which no date has been given.
The licensing matters for companies planning to run the model themselves. Baidu’s English page says the weights will be released under Apache 2.0, a permissive license that allows commercial use, and AI Tech Review reported the same terms. The base model published on November 5 carries a custom ERNIE 5.5 Flash Base Model License Agreement v1.0 instead.
The Preview builds on an open base model from November 5
The base model page describes a sparse mixture-of-experts design, in which only a fraction of the network is used for each word it processes, pretrained from scratch on 33.5 trillion tokens with the Muon optimizer and FP8 mixed precision. Its training context was 128,000 tokens, with fine-tuning and inference supported up to one million. IT Home’s report on the base model noted that Baidu compared it with three similarly sized base models, DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base, run under the same evaluation framework. Base weights are on Hugging Face in FP8 and BF16 versions.
The Preview is the chat-ready version of that checkpoint. Baidu pitches it for coding, 3D modeling, knowledge work and content creation, and trained it on 300,000 knowledge-work workspaces containing 10 million reference files across more than 40 file types, plus more than 150,000 video-creation environments. Baidu says ERNIE 5.5 Flash Preview wrote an entire third-person martial-arts game in C++ for Unreal Engine 5.8 in a single 11-hour coding task. Other demonstrations include nearly 100 mobile mini-games, Blender scenes, FreeCAD models of a W16 engine and a lunar rover concept, a 20-page research report delivered as slides and an Excel model with 1,000 Monte Carlo market paths.
Benchmarks and pricing rest on Baidu’s own figures
Baidu’s pages show benchmark results only as images and describe the scores as "competitive". AI Tech Review read the published chart and reported that across eight benchmarks the Preview ranked first on AutomationBench and SpreadsheetBench V2, both tests of office and automation tasks, and second on DeepSWE, APEX-Agents, DRACO and OfficeQA Pro V2. It warned that comparison figures marked with asterisks are self-reported and should be treated as reference only. No independent evaluation has been published.
The same outlet reported that single-user decoding through the Qianfan API exceeds 200 tokens a second. It listed prices per million tokens of RMB2 for input and RMB8 for output at peak times, and RMB1 and RMB4 off-peak, with a limited-time 50% discount. It also said the open weights will support the SGLang and vLLM serving frameworks.
AI Tech Review placed the release in a week of Ernie activity: the base model on November 5, a first disclosure on November 9 of Baidu’s work on recursive self-improvement through a tool called ERNIE Evolve, and the Preview on November 11. It also noted that Sun Tianxiang, a former lead of the MOSS model project, joined Baidu in July to head its foundation model research department.
The weights are the next checkpoint for Baidu
The commitment Baidu can be held to is the November 25 deadline. Until the weights appear on Hugging Face, the model is available only as a hosted service, and the license that actually ships with the files will decide whether the Apache 2.0 promise on the English page holds.
A second marker is the formal release that Baidu says will follow the Preview. It has not said whether that version will keep the same 280-billion-parameter size or the current Qianfan pricing.