dsh-qwen-multimodal

Text-only main models (e.g. DeepSeek) cannot natively understand images, listen to audio, or generate pictures. dsh-qwen-multimodal fills all three gaps inside a single plugin by calling Alibaba Cloud Bailian Qwen APIs, so the host model can invoke multimodal tools mid-conversation without switching models or restructuring its workflow.

The plugin wraps three core tools: describe_image, powered by Qwen VL, handles image understanding, OCR, and chart analysis with support for multiple images per call; transcribe_audio, powered by Qwen3-ASR, converts wav, mp3, m4a, aac, and other common audio formats into text; and generate_image, powered by Qwen-Image, renders pictures from text prompts and saves them as local files. By design, all media content is processed by the tool layer, and only text transcriptions or file paths return to the main model context, keeping raw pixel and waveform data out of the conversation window.

The generate_image tool also includes a generate-then-verify quality loop: when the verify option is enabled, Qwen VL automatically re-checks the generated result against the original prompt to close the feedback loop. The plugin is ideal for developers working with DeepSeek or other text-only models who need everyday tasks like screenshot OCR, meeting-recording transcription, or quick visual asset generation.