Z.ai releases GLM-TTS, an MIT-licensed Chinese and English voice cloning model

Z.ai published GLM-TTS, a text-to-speech model for Chinese and English, optimised for mixed text. The card describes two stages: a Llama-based language model generates speech tokens, then a flow-matching model produces mel-spectrograms that a vocoder turns into audio. Features include zero-shot voice cloning from 3 to 10 seconds of prompt audio, reinforcement learning (GRPO) with multiple rewards for expressiveness, hybrid phoneme and text input to control pronunciation such as polyphones, and streaming inference. The card does not state a parameter count.

On seed-tts-eval the card reports a character error rate of 1.03 for the base model and 0.89 for GLM-TTS_RL, against 1.12 for the closed-source Seed-TTS and 1.38 for CosyVoice2. Seed-TTS keeps the highest speaker similarity at 79.6, with GLM-TTS_RL at 76.4. These are the vendor's own numbers.

The licence is MIT. The card has no code snippets; quick start means cloning the GitHub repository. The date used is the Hugging Face repo creation date, 10 December 2025; the card says the technical report came out on 16 December 2025.

Lowest error rate in its own table, with the better speaker-similarity score held by a closed model.

Sources

• Z.ai model card on Hugging Face
• Hugging Face repo metadata (creation date)

Editor's first pass (AI-assisted)

• Real 4: MIT weights are on Hugging Face; setup goes through the GitHub repo.
• Useful 3.5: Cloning plus phoneme input helps polyphone-heavy Chinese; two languages only.
• Notable 3.5: Its RL variant tops the card's own character-error table.
• Novel 3: LLM plus flow matching is standard; the multi-reward RL step is the twist.