DeepSeek has introduced its first native multimodal model. The DeepSeek-V4-Flash-Vision-Exp model was made available on the Hugging Face platform on August 31st under the MIT license, marking the initial support for image input in the V4 series alongside text.
The release includes model files, a tokenizer, a minimal implementation for prompt encoding, and a minimal inference implementation in PyTorch. This implementation covers the vision encoder, aligner, DFlash attention, mixture-of-experts routing layer, Hyper-Connections modules, and DSpark.
DeepSeek noted that the model weights are an experimental build, 'Exp'; previously, the model was available via the DeepSeek API starting August 21st. Thanks to its visual capabilities, the model can describe images, read text from screenshots, and analyze charts, accepting inputs in JPEG, PNG, GIF, and WebP formats.
In tasks based solely on text—such as reasoning, agents, and world knowledge—the company asserts that performance matches the V4-Flash version, retaining its previous strengths. However, in agent benchmarks requiring visual understanding, the model significantly surpasses V4-Flash, bringing the capabilities of a multimodal agent closer to the level of Claude Opus 4.8.
By opening up the inference stack and the weights themselves, DeepSeek positions this model for integration into various frameworks and agent tools, expanding its strategy of open weights from text tasks to vision-supported tasks. For developers and enterprises, this release lowers the barrier to entry for locally deploying a competitive model with vision capabilities or using standard infrastructure, aligning with the industry-wide trend toward cost-effective and customizable open models.
Models capable of handling visual data have become key elements in the AI agent market, as developers combine screenshots and real images with reasoning functions. DeepSeek's move follows the trend of increasing open multimodal releases from Chinese labs, which use permissive licenses and aggressive pricing to gain popularity on US developer platforms, cloud service catalogs, and corporate model menus.
Since the model is relatively lightweight for its weight class and comes with a minimal inference reference, teams can quickly integrate it without needing proprietary orchestration tools. The company stated that it will continue to improve the V4 lineup, using the vision-enabled experimental build as a means to test architectural decisions—aligner components, DSpark, and Hyper-Connections—before a broader multimodal release. Currently, the experimental flag indicates the need for caution: chart recognition and understanding are strong, but DeepSeek advises testing the model on specific workloads rather than assuming production readiness across all image tasks. Early users, including research groups and application developers, are using the open stack to assess how well the multimodal agent's Opus-level skills hold up in their own processes.
