Dots Studio: Dots3-Note Preview (free)

dots3-note Preview is Dots Studio’s lightweight member of the dots3 multimodal family. The producer lists a mixture-of-experts model with 280B total parameters, 16B active parameters and support for up to 512K tokens of context. It takes text, images, video and audio and produces text, with general knowledge, instructions, reasoning, coding and agent-oriented work among its stated tasks.

The language backbone has one dense layer and 45 expert layers, with eight routed experts selected from 256 alongside a shared expert. Its attention combines 13 sparse-attention layers and 33 sliding-window layers. The producer lists separate vision and audio encoders and a shared prediction layer. These components support multimodal understanding; the model’s output remains text rather than generated images, speech or video.

For visual analysis, supply the image, page or chart and ask about the feature that affects the decision. A useful document task can compare a visible table with a written requirement and return the relevant row, the interpretation and any missing information. For video, identify the event or sequence to examine; the producer’s input example includes the audio track when the clip has one.

For audio, state whether the task needs a transcript or an explanation of the material, and provide any names or domain terms needed for interpretation. A combined assignment can use the recording together with written background and ask for findings organized by the question they answer. The result is a textual artifact for the application to present or process.

The producer’s conversation template distinguishes thinking from a direct response. Enabling thinking adds a reasoning stage; disabling it returns the answer directly. The documented sample uses temperature 1.0 and top-p 0.95. Choose the task and desired result explicitly, and leave enough generation room for the answer rather than treating a sample token allocation as the model’s hard output limit.

For a coding task, provide the relevant source, the expected behavior and an observed failure or requested change. Ask for the proposed artifact and the checks that would assess it. In a connected agent workflow, the model proposes tool interactions and the application executes them and returns results. The maker publishes a dedicated tool-call parser; the surrounding application supplies the actual tools and their access.

Interactive work can require exploration, adaptation and updates to the task’s working memory. Keep a compact record of the objective, confirmed findings and unresolved condition. Ask the next turn to identify what new evidence changes and what should remain in that record. Persistence of the record and execution of actions belong to the application around the model.

The producer publishes distinct BF16 and FP8 artifacts and local serving recipes with different configured context lengths. A configured serving length is separate from the 512K model specification. Preserve the Preview edition and precision distinctions when identifying the model, and request a final result that states the finding, the proposed change and the evidence the workflow actually gathered.