dots3-note Preview goes open-weight, multimodal

This is a serious open-model push toward agentic workloads, pairing frontier-scale capacity with a comparatively small active footprint. Its real test will be whether the multimodal and long-context claims translate into practical inference costs and reliable tool use.

• –16B active parameters could make a 280B model more deployable than its headline size suggests.
• –512K context is valuable for codebases, research archives, documents, and persistent agent sessions.
• –Native image, video, and audio understanding broadens use beyond conventional text-only agents.
• –Open weights give developers room to self-host, fine-tune, and benchmark the model independently.
• –Preview status means inference support, evaluation quality, and production stability remain open questions.