多模態模型與 API
DeepSeek V4-Flash-Vision-Exp Adds Image Input and a 1M-Token Context Window, but Remains an Experimental API
DeepSeek has added image understanding to V4-Flash, supporting text and images through three compatible interfaces while retaining the existing Flash pricing. However, its officially reported agent benchmark results have yet to be independently reproduced, and neither the model weights nor the vision architecture has been disclosed.
On August 21, DeepSeek launched `deepseek-v4-flash-vision-exp` on its official API, adding image input to the V4-Flash family, which had previously focused primarily on text and coding-agent tasks. The model accepts JPEG, PNG, GIF, and WebP files. Developers can encode images as Base64, provide public URLs, or upload them through the Files API before submitting them through the Chat Completions, Messages, or Responses interface. Output remains text-only.
The specification page lists a 1 million-token context window and a maximum output of 384,000 tokens. Images are resized and converted into visual tokens, with each image using up to approximately 384 tokens. A low-resolution mode is available for agents that require only coarse visual recognition. Images may appear only in user messages; existing applications that attach files to the system or assistant role will receive a request error. These constraints mean that “OpenAI-compatible” does not guarantee plug-and-play support in every existing client: model catalogs and provider adapters must also explicitly declare the input modalities as `text, image`.
DeepSeek reports scores of 83.9 on Terminal-Bench 2.1, 57.7 on NL2Repo, and 59.3 on DeepSWE, while claiming that text-only performance is comparable to V4-Flash. The evaluations used the DeepSeek Harness in minimal mode, the highest reasoning effort, and specified sampling settings. The scores therefore reflect both the model and the agent scaffolding and should not be treated as measurements of the standalone model’s capabilities. Several of the benchmarks are also not purely vision-focused tests.
The engineering appeal is that a single agent can process terminal screenshots, charts, and document pages without requiring a separate OCR pipeline or another VLM. The model still carries the `-exp` suffix and is available only through the API. DeepSeek has not released the weights, vision encoder, projection layer, training recipe, or image-token layout. Teams should first evaluate accuracy on real documents containing small text, dense tables, and multiple ordered images, maintain a fallback route to a text model, and avoid treating DeepSeek’s self-reported scores as a guarantee of production stability.