Seed 1.6 (250915)

Seed 1.6 comes from ByteDance's Seed family of vision models, released in September 2025 as a multimodal large language model capable of processing both text and image inputs. What sets this model apart is its dual-mode design offering both thinking and non-thinking response styles, allowing users to toggle between explicit chain-of-thought reasoning and faster direct responses depending on task complexity. This flexibility suggests the model was designed to handle everything from deep analytical problems requiring step-by-step reasoning to simpler queries where responsiveness matters more.

The model traces its lineage to ByteDance's broader Seed development effort, with the September 2025 release marking a stable, production-ready vision capability. The thinking/non-thinking architecture points to deliberate post-training refinement, likely involving techniques that cultivate reasoning flexibility while maintaining broad task coverage. Being available as open weights and served across multiple providers including OpenRouter and Vercel AI Gateway reflects strong ecosystem demand for adaptable vision models. With a 256,000-token context window, the model is well-suited for complex document understanding, visual question answering, and applications where the reasoning mode can be matched to problem difficulty.