GitHub radar
DeepSeek Flash can now view images!
DeepSeek has launched an experimental multimodal variant of their Flash model, allowing it to process images and text. This allows them to create vision-enabled agents cheaply.
DeepSeek-V4-Flash-Vision-Exp is DeepSeek’s first experimental multimodal model in their DeepSeek-V4 series. By taking their already impressive DeepSeek-V4-Flash model (which excelled at quickly and affordably performing agentic tasks), they have equipped it with vision capabilities. Now, their model is able to process both images (such as screenshots, flowcharts, graphs, and mockups) as well as text. As a result, DeepSeek reports that their multimodal agent abilities have seen a major boost, without sacrificing any text-based functionality compared to the original Flash model. This is an experimental checkpoint, and will be updated in future releases.
Why a vibe-coder should care
With vision-enabled agents, you’ll be able to describe a problem via a screenshot or mockup, instead of needing to write out a detailed explanation. Since the DeepSeek Flash model is much more affordable than other multimodal models, this gives you access to GPT-4o level vision tasks for a much lower price.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Connect to DeepSeek-V4-Flash-Vision-Exp via API (https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-Vision-Exp) and show how to send it an image alongside a text question and get a response — for example, a screenshot description.
Server-grade model — 305 billion parameters, you can't run it at home. Available through the DeepSeek API.
Open on Hugging Face▌ More finds