DeepSeek V4 Flash adds a multimodal Vision API to its agentic deployment options

DeepSeek has added DeepSeek-V4-Flash-Vision-Exp to its API platform as an experimental multimodal companion to V4 Flash. The new model keeps DeepSeek’s claimed text capabilities at V4 Flash level while accepting mixed text-and-image input through Chat Completions, Messages and Responses. Images can be supplied as base64, external URLs or Files API references, are tokenized at up to 384 tokens each, and are billed at the same V4 Flash rates. DeepSeek also launched a free Files API so an image can be uploaded once and reused by file_id. Separately, the official DeepSeek-V4-Flash-0731 text weights remain MIT-licensed with the existing 1M context, 304B-parameter footprint and documented vLLM/SGLang self-hosting path.

The practical choice around V4 Flash is no longer just hosted-versus-self-hosted text inference. Builders with browser automation, document understanding, visual QA or multimodal agent workflows can now test a low-cost DeepSeek vision path without changing away from the V4 Flash family, while teams that need private text inference still retain the open-weight deployment option. The distinction matters because the Vision-Exp release is an API product, not an announced open-weight checkpoint, so portability and data-control assumptions differ between the two paths.

DeepSeek-V4-Flash-0731 remains the official downloadable text model under MIT, with roughly 304B parameters and documented vLLM and SGLang serving. DeepSeek-V4-Flash-Vision-Exp is a separate experimental API model. DeepSeek has not announced downloadable Vision-Exp weights, so builders should not assume the self-hosting rights and deployment path of 0731 automatically extend to the multimodal variant.

Vision-Exp supports mixed text and image input through OpenAI-style Chat Completions and Responses as well as DeepSeek’s Anthropic-compatible Messages path. Images can be passed inline, by external URL or through the new Files API. Uploading once and reusing a file_id can reduce repeated request payload size for agents that inspect the same screenshots, diagrams or documents across multiple turns.

DeepSeek says each image is tokenized at up to 384 tokens and billed at V4 Flash pricing. That keeps visual input costs predictable relative to the existing text model, although total task cost still depends on generated text, reasoning depth and repeated image use. Builders should measure end-to-end agent cost rather than comparing only the per-image token ceiling.

DeepSeek says Vision-Exp matches V4 Flash on text capabilities and makes a large jump on multimodal-agent benchmarks, approaching Opus-4.8 on its published comparisons. Those results use DeepSeek’s own harness and settings and include internal or harness-sensitive tests. They are useful as a reason to evaluate the model, not as a substitute for workload-specific testing.

The existing V4 Flash 0731 weights preserve a private-inference path for coding and long-context text workloads. The new vision capability currently changes the hosted-service option instead. Teams with regulated images, proprietary screenshots or customer documents should therefore separate text-model openness from multimodal data-handling requirements when choosing an architecture.