ByteDance-Seed is a research team established in 2023 within the Chinese technology conglomerate ByteDance Ltd. The team was formed with the ambitious mission to “discover new approaches to general intelligence.” Operating as a multi-site laboratory with centers in China, Singapore, and the United States, the “Seed” brand serves as an umbrella label for all of its scientific publications and model releases. While ByteDance’s historical journey in AI began with the development of recommendation systems for its Jinri Toutiao platform, the creation of Seed in 2023 marked the company’s strategic pivot toward fundamental research in large language and multimodal models aimed at achieving AGI.
ByteDance-Seed’s major LLM releases occurred only in 2025. In April, the team unveiled Seed1.5-Thinking — a model introducing the concept of “controllable thinking budget,” where the model explicitly generates structured reasoning traces before producing a final response. By May, the team extended this paradigm into the multimodal domain with Seed1.5-VL. This model achieved state-of-the-art results on 38 out of 60 public vision-language benchmarks. Its engineering elegance lies in the use of a compact visual encoder (only 532 million parameters) combined with a sparse MoE architecture, enabling high performance and strong agentic capabilities such as GUI control. Culminating this research cycle, in August the team released Seed-OSS-36B — a family of fully open models with 36 billion parameters. Here, Seed implemented its cutting-edge innovations in a format accessible to the global research community. Notably, it became the industry’s first model to offer a native context window of 512,000 tokens — engineered directly during pretraining without relying on external retrieval or interpolation techniques. These models are positioned as “research-friendly,” optimized for tasks requiring long-context processing, complex reasoning, and agentic behavior.
ByteDance-Seed has actively integrated itself into the global AI industry. Rather than simply releasing model weights, the team first establishes methodological foundations (e.g., benchmarks and code sandboxes), then publishes scientific reports, and only afterward releases open models with well-documented innovations — such as native ultra-long context and controllable reasoning. These advancements are rapidly integrated into ByteDance’s commercial products, most notably Doubao, an exceptionally popular chat and inference platform in China.