Taobao Flash Sale releases and open-sources the food safety governance AI model "Bai Ze"

Shanghai Securities News China Securities Network News (Reporter Yang Xiangfei): The food safety governance field in the catering industry has gained a powerful new AI tool. Recently, Taobao Flash Sale officially released an open-source large model—“Bai Ze” (Ostrakon-VL)—designed specifically for food service and retail stores, and announced that the related technological capabilities are now freely available to the entire industry.

“Bai Ze” is now open source on Hugging Face

Currently, “Bai Ze” has been integrated into over 100 production scenarios on Taobao Flash Sale, with call volumes exceeding 1 billion, and token consumption surpassing 10 trillion. More importantly, the deployment cost of “Bai Ze” is only 1/20 of similar solutions, helping industry participants in online catering to implement large-scale applications at very low costs.

With the help of the “Bai Ze” large model, industry participants such as delivery platforms, catering, and retail companies can quickly enhance their ability to identify and manage non-compliant scenarios through AI technology in areas like image recognition and kitchen alerts, significantly improving digital governance efficiency.

As the first multimodal large language model tailored for food service and retail store scenarios, “Bai Ze” has been actively involved in the daily platform governance of Taobao Flash Sale, playing a key role in core scenarios such as merchant onboarding review, daily inspections, and risk alerts. Taobao Flash Sale will continue to expand the application scope and depth of “Bai Ze,” focusing on tackling persistent food safety issues that trouble the industry.

Currently, Taobao Flash Sale supports merchants uploading “one-shot” videos to label their stores as “dine-in available.” “Bai Ze” can also analyze the video content and store-related information to comprehensively determine whether the store has a genuine dine-in environment, using cross-frame consistency analysis to verify the authenticity of videos, preventing the use of static images or AI-generated videos to deceive, and ensuring consumers receive authentic service experiences.

夏威, the project leader of “Bai Ze,” explained that the model is based on Alibaba Cloud’s Tongyi Qwen3-VL-8B foundation, which has been deeply trained and fine-tuned to meet the specific needs of food and retail store scenarios. Its main technical advantage is extremely high parameter efficiency. According to the ShopBench evaluation (a public benchmark for catering and retail stores), “Bai Ze” significantly outperforms models of the same scale in performance scores, maintaining robust perception and decision-making abilities even in noisy, highly heterogeneous real store data.

A relevant person in charge of Taobao Flash Sale technology stated that the platform has long been increasing investment in AI research and development, using technological innovation to safeguard “the safety on the tip of the tongue,” ensuring consumers can shop with confidence and providing compliant merchants with a more favorable operating environment.