docs: clarify Zen fine-tune of Qwen3 audio base + honest base_model attribution

README.md

@@ -5,27 +5,36 @@ pipeline_tag: text-to-speech
 language:
 - en
 - zh
+base_model:
+- Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign
+base_model_relation: finetune
 tags:
 - zen3
 - zen3-tts
 - zenlm
-- hanzo
+- qwen3-tts
 - speech-synthesis
 - voice-design
 - controllable
 - 12hz
-- zen-3
 ---
 
 # Zen3 TTS Voice Design
 
-Zen3 TTS variant specialized for voice design: synthesize new speakers from natural-language descriptors (timbre, age, accent, affect) without reference audio. Part of the Zen3 audio family.
+Zen3 TTS variant for voice design: synthesize new speakers from natural-language descriptors (timbre, age, accent, affect) without reference audio.
 
-## Hosted via Hanzo gateway
+Derived by fine-tuning [`Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign) (Alibaba Cloud, Apache-2.0).
 
-Served at `api.hanzo.ai` as `zen3-tts-voice-design`.
+- Architecture: `Qwen3TTSForConditionalGeneration` (`qwen3_tts`)
+- Parameters: ~1.7B
+- Base model: [`Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign)
 
 ## Weights
 
-Mirror staging in progress. Use the canonical base from the Zen model card until weights land here.
+This repository contains the model weights: `model.safetensors` (talker) plus a `speech_tokenizer/` module (12 Hz codec), config and tokenizer files.
 
+The model uses the `qwen3_tts` architecture and loads with `transformers` (>= 4.57). It is API-compatible with the upstream base — follow the inference recipe on the base model card [`Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign).
+
+## Provenance
+
+Fine-tuned from [`Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign`](https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-VoiceDesign) (Apache-2.0). See `NOTICE` for full attribution.