Copyright 2026 Adventists.ai.

Sophon-Realtime-32B-Tag bundles model weights derived from the following works, each licensed under the Apache License 2.0:
- thinker/: derived from Qwen3-VL-32B-Instruct by the Qwen team, Alibaba (https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct),
  via DuplexJev-32B-Para (Adventists.ai) with a merged tag-writing adapter; quantized to FP8. Its audio encoder is derived from
  Qwen3-ASR (Qwen team, Alibaba, Apache License 2.0).
- talker/: tag-tts (Adventists.ai), derived from Fun-CosyVoice3-0.5B by the FunAudioLLM team, Alibaba
  (https://github.com/FunAudioLLM/CosyVoice); the language model was fine-tuned to perform sound tags, the flow model was
  distilled to 4 steps, and the HiFT vocoder is redistributed unchanged.
