LongCat-AudioDiT lands with waveform latent TTS

This is a meaningful speech-model release because it attacks TTS complexity at the representation level, not just by scaling up another pipeline.

• –Direct waveform-latent generation removes the mel bottleneck and should reduce compounding errors in long-form synthesis.
• –The APG guidance swap and training-inference mismatch fix look like the real quality wins, not just parameter count.
• –The 3.5B model’s Seed gains are modest but credible, especially for zero-shot cloning where speaker similarity matters a lot.
• –The paper’s Wav-VAE finding is important: better reconstructions do not automatically translate to better end-to-end TTS.
• –Open weights plus a Hugging Face-compatible implementation make this immediately useful for downstream speech tooling and fine-tuning experiments.