ByteDance's Seed team has officially launched Seed 2.0, a new series of AI models designed for large-scale production deployment. The series includes three Agent models: Pro, Lite, and Mini, along with a Code model. The Pro and Code models are already available on the Doubao App and TRAE, and the full series is accessible via API on Volcano Engine.
According to the announcement, Seed 2.0 represents a step up from its predecessor, advancing from Olympiad-level to research-level reasoning. The models are said to handle hour-long videos efficiently, with a VideoCut tool that extends processing duration for even longer content. They also show improved visual perception, long-context understanding, and document comprehension for complex layouts.
On benchmark tests, Seed 2.0 Pro scored higher than GPT-5.2 on SuperGPQA and achieved gold-medal-level performance on competitive programming and math exams like ICPC, IMO, and CMO. The company reports that Seed ranks 6th on the LMSYS Chatbot Arena Text Arena Overall and 3rd on the LMSYS Vision Arena leaderboard, as of February 16, 2026.
The company says it tuned the series using service usage data from the "Volcano Ark Collaboration Incentive Program" in China's mainland, gathered from users who signed authorization agreements. The models are designed to prioritize user experience under large-scale online deployment, supporting tasks with multiple constraints and steps. The announcement also describes a shift from passive question and answer exchanges to proactive guidance in areas such as fitness and styling.
The Seed 2.0 homepage includes a model card at seed.bytedance.com/seed2. In a separate post, the Seed team introduced a full-duplex speech model that it says allows more natural interaction.
