DeepSeek adds vision to V4 Flash at experimental API launch

DeepSeek has added image understanding to its V4 Flash line with deepseek-v4-flash-vision-exp, an experimental model now available through the company's API. It accepts text and images, returns text, and keeps the same listed per-token rates as V4 Flash. The release gives developers a low-cost way to test visual inputs in long-context agents, but it does not come with open weights or an independent benchmark evaluation.

What DeepSeek shipped

The new model is exposed as deepseek-v4-flash-vision-exp. DeepSeek's API documentation lists support for Chat Completions, the Responses API, JSON output, tool calls, and its Anthropic-compatible API. The model has a 1M-token context window and a maximum output of 384K tokens. Images are converted into input tokens according to their dimensions and billed with the rest of the request.

The documentation describes the release as experimental. That matters for production planning: the model is an API endpoint, not an open-weight publication, and DeepSeek says product prices can change. The current page also shows a concurrency limit of 2,500.

Pricing stays in the V4 Flash lane

DeepSeek's current pricing table places the vision model in the same listed rate band as V4 Flash. Peak weekday hours—01:00–04:00 and 06:00–10:00 UTC—are billed at higher rates, and images are converted into input tokens according to their dimensions.

There is no separate image surcharge in the table. Image size still affects the number of input tokens, so a visual workload's cost depends on the files and the surrounding prompt, not only on the model name.

The benchmark claim needs a narrow reading

DeepSeek says the model preserves V4 Flash's text capabilities and makes a major improvement on multimodal agent tasks. The company compares it with Anthropic's Opus 4.8 in an eleven-benchmark table. The Next Web reports that DeepSeek wins three rows, trails on the other eight, and falls twelve points behind on NL2Repo.

That comparison is useful evidence of the vendor's target, not a neutral verdict. DeepSeek ran the evaluations with its own harness and disclosed that the text-only baseline ignores visual elements in two multimodal tests. The independent report also notes that the table does not compare the new model with Anthropic's newer Opus 5. Until outside evaluators reproduce the results, the practical claim is simpler: DeepSeek has shipped a vision-capable API model whose cost profile is attractive enough to test, while its relative quality remains unsettled.

How developers can evaluate it

Start with workloads where visual context is central: screenshot interpretation, document images, UI-state extraction, and agents that need to inspect a screen before calling tools. Compare it with the text-only V4 Flash model on the same prompts, images, tool definitions, latency budget, and failure-handling logic. Record image-token usage separately from text-token usage.

For production use, verify retention, rate limits, regional requirements, and model lifecycle directly against DeepSeek's current documentation. Keep a fallback model available while the endpoint carries the experimental label.

Sources and methodology

This article uses DeepSeek's API model documentation and pricing page as primary sources, with The Next Web as independent context. DeepSeek's benchmark figures are presented as vendor-reported results; no independent reproduction is implied.

For more practical model and agent workflows, see LinkLoot's AI agent tools guide.