DeepSeek unveils new multimodal model with vision capabilities

On August 31, 2026, DeepSeek released its first multimodal model, DeepSeek-V4-Flash-Vision-Exp, on Hugging Face under the MIT License. The model includes files, tokenizer, prompt encoding, and a minimal PyTorch inference implementation covering core modules like vision encoder, Aligner, DFlash Attention, and MoE. DeepSeek-V4-Flash-Vision-Exp, marked as an experimental version, was launched on August 21, 2026, and supports image input in addition to text. It can describe images, recognize text in screenshots, and analyze charts in JPEG, PNG, GIF, and WebP formats. DeepSeek claims the model performs well in multimodal agent frameworks, matching the V4-Flash formal version in text-based tasks and outperforming it in visual understanding benchmarks, nearing the capabilities of Opus-4.8. The model's vision capabilities allow it to handle tasks requiring image analysis, expanding its utility in various applications. DeepSeek emphasizes that the model retains the strengths of the V4-Flash version while significantly improving its performance in visual tasks, making it a strong contender in the multimodal AI space.