Qwen3.8 27B dense vision-language model on self-hosted GPUs. Text and image input (OpenAI image_url, base64 data URL or https URL), streaming. Reasoning...