Dots3 opens a 280B multimodal model

Dots Studio has released dots3-note Preview, an Apache-licensed multimodal mixture-of-experts model with 280 billion total parameters, 16 billion active parameters, and a 512K-token context window. The model card says it accepts text, images, video, and audio, while the official repository recommends serving its FP8 checkpoint on one eight-GPU node; vLLM support is on the main branch while Transformers and SGLang integrations remain under review. Local-model teams should file the release as a broad capability-per-rack option, not a frictionless deployment: multimodal runtime maturity and eight-GPU economics still matter alongside the archive’s open-weights control analysis.