News Summary

• DeepSeek-V4-Flash-Vision-Exp is our first experimental multimodal model.
• It builds on the DeepSeeks architecture by incorporating visual modules.
• This repository contains the tokenizer, prompt encoding reference, and a minimal PyTorch inference implementation.
• The reference inference covers the vision encoder and aligner, DFlash attention, MoE, Hyper-Connections, and the DSpark forward path.
• See the SGLang cookbook for detailed instructions, benchmarks and other hardwares configurations.