DeepSeek posted documentation for deepseek-v4-flash-vision-exp, an experimental multimodal version of its low-cost V4 Flash model. The key implementation detail is aggressive image normalization. Every image is resized to roughly 800×800 total pixels before inference, with a cap of 384 image tokens, which makes pricing predictable and cheap but immediately constrains high-detail use cases. That tradeoff shaped almost all of the reaction.
The excitement came from a very practical place. A lot of people already use V4 Flash as a coding model and have been missing vision for the last mile of agent workflows. The new model can now inspect screenshots, check rendered UIs, look at charts and sketches, and close the feedback loop on browser or desktop automation without handing off to another model. Several people said that was the missing piece keeping DeepSeek from being a stronger daily driver for frontend work, game development, and visual QA.
The main skepticism was not about whether vision is useful. It was about whether 800×800 is enough. People doing
OCR on full documents, reading small text, parsing sheet music, understanding diagrams, or recognizing landmarks found the model shaky or outright bad. The clock test became a proxy for that broader concern. Some saw it as a silly gotcha. Others treated it as a clean check for jagged intelligence in a model that claims general visual understanding. The more useful conclusion was that this release is good at screenshot-scale agent tasks, not a drop-in replacement for high-fidelity vision systems.
A strong practical theme ran through the discussion: much of the limitation can be pushed into the harness. People described multi-turn crop tools, grid tiling, local preprocessing, and document-layout models that split pages before sending them to the API. That does not magically solve all spatial reasoning failures, but it does move this model from “too low resolution” to “good enough for many workflows” when the task is inspectable in parts. Comments also clarified that image outputs from tool calls are possible if you use APIs that support them, which matters for screenshot-checking agents.
The overall read was upbeat but unsentimental. DeepSeek finally added the capability many users wanted most, and the price profile looks attractive for software engineering agents. At the same time, nobody serious came away thinking this is frontier-grade vision across the board. It looks like a cheap, useful multimodal coding model with obvious blind spots, and a lot of the value depends on how much scaffolding you are willing to build around it.
api-docs.deepseek.com
Clock test image 1
Clock test image 2