The release adds image input to a model whose text capabilities the company says match V4-Flash. It supports mixed text and image input via base64, external URLs, or the Files API, and works through Chat Completions, Messages, and Responses. DeepSeek also says Harness 0.1.1 includes out-of-the-box support.
The thesis is: This release makes visual input a callable extension of DeepSeek’s existing Flash API surface, rather than a separate one-off interface. The mechanism is: A shared model endpoint accepts text and images, with image billing capped at 384 tokens per image at V4-Flash pricing; the Files API adds free upload and reuse through file_id. This can reduce integration friction for teams whose agent workflows need to inspect images, though the source does not document task-level reliability or deployment outcomes.
Why now: DeepSeek’s announcement moves the V4-Flash line beyond text-only use and names supported interfaces and input paths, giving builders something concrete to test. The first implication is: Teams already using these API interfaces can evaluate visual tasks without switching to a separate endpoint family. The first watch condition is: Whether independent evaluations and user experience establish reliable performance on real multimodal agent tasks. Evidence boundary: The “close to Opus-4.8” benchmark comparison is DeepSeek’s characterization; the announcement provides no benchmark methodology or independent validation. This begins a distinct analytical record because no prior New Runtime record was supplied for this specific vision API release.