DeepSeek V4 Flash Vision

An experimental vision build of V4 Flash: the same sparse Mixture-of-Experts design with 13B active parameters out of 284B total, extended to read images while matching the base model on text, agents, reasoning and world knowledge. DeepSeek aims it at document and chart understanding, visual question answering and multimodal agent workflows that interleave text and images, and reports its multimodal agent scores as a large jump over plain V4 Flash. Images are tokenised for billing at up to 384 tokens each, so vision costs V4 Flash rates. It is labelled experimental by DeepSeek, and the id may change.

This is V4 Flash with eyes, and DeepSeek is careful to say it is nothing more than that on the text side. It is the same sparse Mixture-of-Experts design - 13B active parameters out of 284B total - and the lab's claim is that it matches plain V4 Flash on text capabilities including agents, reasoning and world knowledge, with image understanding added rather than traded for.

Where it moves is the multimodal half. DeepSeek reports that on multimodal agent benchmarks the model makes a major leap over V4 Flash, bringing multimodal agent performance close to Claude Opus 4.8. That is the claim the release exists to make, and it is a specific one: not that the model sees well in isolation, but that it holds up when seeing is one step inside a longer agentic loop.

The uses DeepSeek names follow from that framing - document and chart understanding, visual question answering, and multimodal agent workflows that interleave text and images. It is the interleaving that matters. A model that reads a chart is useful; a model that reads a chart in the middle of a tool-calling sequence, without losing the thread of what it was doing, is a different capability, and it is the one being claimed here.

Billing is the part most likely to surprise. Images are tokenised at up to 384 tokens each and charged at ordinary V4 Flash rates, so vision costs essentially nothing beyond the tokens - a screenshot is priced like a short paragraph. Access to the Files API is free. Against a model already among the cheapest capable options available, that makes visual input unusually inexpensive to use in volume.

The caveat is in the name and DeepSeek does not soften it: this is labelled experimental, and its id carries an "exp" suffix to say so. Experimental ids get renamed, superseded or withdrawn on the lab's schedule rather than yours, so treat it as a capability to try rather than one to build a dependency on. There is no separate published architecture or context detail beyond what V4 Flash already carries.