ByteDance announces Seed 2 multimodal LLM

After making waves with its Seed Dance 2.0 video generator, ByteDance has released Seed 2, a multimodal large language model aimed squarely at frontier-class competitors. Benchmarks position Seed 2 Pro alongside Claude, GPT 5.2, and Gemini 3 on a range of tasks spanning reasoning, perception, and agentic behavior.

On visual reasoning and motion understanding, Seed 2 reportedly outperforms several closed models in many tests, though its instruction-following and real-world-task performance are more comparable than dominant. In one demo, the model converts a static chart image into Python code that recreates the plot; in another, it autonomously drives a CAD application through nearly 100 sequential steps, manipulating 3D objects via clicks and UI interactions.

Independent leaderboards like LMSys’s arenas already include Seed 2.0, where it ranks around fifth for text-only tasks and fourth for vision, trailing the latest Gemini entries but ahead of many peers. Access is currently gated through Volcano Engine’s API, which serves as ByteDance’s cloud platform for model deployment.