Seed2.0 has officially released its latest model series, offering three general-purpose agent models—Pro, Lite, and Mini—designed for complex real-world tasks at production scale.

Three Models, One Goal

The Pro model focuses on long-chain reasoning and robustness in complex workflows, optimized for high-stakes scenarios. Lite and Mini round out the series for lighter deployments, though specific distinctions between the three were not detailed in the announcement.

Across all variants, Seed2.0 delivers a comprehensive upgrade in multimodal understanding, with strengthened LLM and agent capabilities that enable steady progression in long-horizon tasks. The company positions the series as expanding from competition-level reasoning into research-grade workloads, claiming first-tier industry performance on high-economic-value and high-scientific-value evaluations.

Visual Understanding and Interaction

Seed2.0 demonstrates significant enhancements in visual reasoning and perception, achieving state-of-the-art performance on benchmarks such as MathVision. For dynamic scenarios, the model strengthens its understanding of temporal sequences and motion perception, securing a leading position on MotionBench.

The announcement highlights several interactive applications:

· Image to web page: The model can turn a single image into a functional front-end web page, autonomously restoring layout and visual details while generating code that supports animation and interactive extensions.

· Real-time fitness assistant: Seed2.0 analyzes user movement data in real time to provide personalized fitness guidance.

· Chart understanding and reconstruction: Based on an input chart image, the model generates accurate matplotlib Python plotting code, effectively converting visual representations back into executable code.

Professional Task Execution

Seed2.0 maintains high stability and reliability when executing multi-step, long-horizon instructions. Examples provided include:

· CAD design and operation: End-to-end modeling of double boss features and extraction of geometric parameters using the FreeCAD Part Design Workbench.

· Biotechnology research support: Drafting well-structured, scientifically sound experimental solutions.

· Algorithm repair: The Solovay-Kitaev Algorithm repair task, which requires strong numerical analysis and geometric algebra knowledge, demonstrates the model's mathematical reasoning and code repair capabilities.

Evaluation Results

The announcement states that Seed2.0 shows comprehensive improvements over its predecessor Seed1.8 across LLM, VLM, and agent task evaluations, with particular strength in reasoning, complex instruction execution, and multimodal understanding.