DeepSeek V4 Flash Vision emerges as an open-weight multimodal agent model

DeepSeek-V4-Flash-Vision-Exp is reported on Hugging Face as a 305B-parameter MIT-licensed model with inference code and multimodal-agent benchmarks. Its native vision path reportedly improves visual tool use while preserving text-agent capability, but absent vLLM support, unavailable hosted inference, and company-reported comparisons make deployment and performance claims provisional.