DeepSeek ships an experimental vision-enabled variant of V4 Flash

DeepSeek has released DeepSeek-V4-Flash-Vision-Exp, an experimental multimodal version of its V4 Flash model that adds image understanding on top of the model's existing text capabilities.

What's new

The new variant accepts images alongside text — via base64 encoding, external URLs, or DeepSeek's Files API — across the Chat Completions, Messages, and Responses APIs. It's live now on the DeepSeek API Platform under the model name deepseek-v4-flash-vision-exp. Images are tokenized at up to 384 tokens each, and DeepSeek is pricing the model the same as the text-only V4-Flash.

On text-only tasks — agentic workflows, reasoning, and world knowledge — DeepSeek says the vision variant matches the base V4-Flash model. The image-related jump is on multimodal agent benchmarks, where DeepSeek describes the improvement over V4-Flash as substantial: "On multimodal agent benchmarks, V4-Flash-Vision-Exp makes a major leap over V4-Flash, bringing multimodal agent performance close to Opus-4.8."

Context

V4-Flash-Vision-Exp builds on DeepSeek-V4-Flash-0731, the sparse mixture-of-experts model (284B total parameters, 13B active) that DeepSeek re-post-trained and published in July. Rather than releasing a new flagship, DeepSeek is iterating on its efficient Flash line, adding modalities incrementally while keeping the same MoE backbone and pricing.

The "Exp" designation signals this is an experimental release rather than a fully supported production model, consistent with DeepSeek's pattern of shipping capability previews on its API before formal, non-experimental launches.

Why it matters

DeepSeek's V4-Flash line competes primarily on price and efficiency rather than raw benchmark supremacy, and extending that line to multimodal agent tasks — while holding text performance and pricing steady — keeps that positioning intact for developers building document-understanding, chart-reading, and screenshot-analysis agents on a budget. Closing the gap to Opus-4.8 on multimodal agent benchmarks, even as an experimental release, is a signal that the vision-agent gap between low-cost and frontier models is narrowing faster than pricing alone would suggest.