ByteDance Seed released UI-TARS, an open-weight native GUI agent model that perceives raw desktop and mobile screenshots to execute mouse and keyboard actions without relying on brittle DOM or accessibility trees.
ByteDance Seed (Seedance & Seedream)
Also known as: ByteDance Seed · Seedance · Seedance 1.0 · Seedream · Seedream 3.0 · BAGEL · UI-TARS
Plain English
ByteDance's frontier AI research division—creator of the top-ranked Seedance video generation models, Seedream image models, the open-weight BAGEL unified multimodal model, and the UI-TARS computer-use agent.
Technical Overview
ByteDance Seed builds foundational models across text, vision, and video: Seedream 3.0 (native high-resolution bilingual image DiT), Seedance 1.0 (decoupled spatial/temporal video DiT with native multi-shot narrative transitions and RLHF motion alignment), BAGEL (Apache-2.0 Mixture-of-Transformer-Experts unifying VLM understanding and visual generation), and UI-TARS (native GUI agent).
Timeline (4)
2025
Paper- Paper
ByteDance Seed published Seedream 3.0, a bilingual text-to-image DiT trained natively at up to 2K resolution with cross-modality RoPE and defect-aware human-feedback reward modeling for dense typography and photorealistic portraits.
- Release
ByteDance Seed open-sourced BAGEL (7B active / 14B total parameters) under Apache-2.0, using a Mixture-of-Transformer-Experts architecture to combine vision-language understanding, text-to-image synthesis, and free-form world-knowledge image editing in a single model.
- Paper
ByteDance Seed introduced Seedance 1.0, pairing decoupled spatial and temporal DiT layers with multi-shot narrative captioning and multi-dimensional video RLHF to generate 1080p multi-cut video clips with high motion stability.
See also (8)
Open full graphByteDance Seed (Seedance & Seedream)