DeepSeek V4's first vision model is now open source, with multimodal agent capabilities close to Opus-4.8

DeepSeek released the first multimodal model of the V4 series, DeepSeek-V4-Flash-Vision-Exp, on Hugging Face on August 31, open-sourced under the MIT License. The release includes model files, tokenizer, reference implementations for prompt encoding, and a minimal PyTorch inference implementation, covering core modules such as vision encoder, Aligner, DFlash Attention, MoE, Hyper-Connections, and DSpark.

This experimental version supports image input, enabling image description, text recognition in screenshots, and chart analysis, with support for JPEG, PNG, GIF, and WebP formats. The company says it matches the official V4-Flash release on pure text tasks like agent, reasoning, and world knowledge, while showing significant improvement on agent benchmarks requiring visual understanding, bringing its multimodal agent capabilities close to Opus-4.8.