Fish Audio has introduced Drama 3, an open-weight text-to-speech model featuring 35 billion parameters. The model utilizes a mixture-of-experts architecture based on the Qwen3.5-35B-A3B backbone. It supports 83 languages and allows for up to 16 distinct speakers in a single API request.
The technical architecture incorporates a new Codebook-MTP decoder and utilizes the Descript Audio Codec to generate 44.1 kHz audio. Training involved three pre-training stages, followed by instruction tuning and reinforcement learning using GRPO. The model achieved a 93.2% score on the Artificial Analysis Pronunciation Robustness benchmark.
Deployment requires at least 48 GB of NVIDIA GPU memory in FP8 format, with no support for CPU, Apple MPS, or AMD ROCm. On an H200 GPU, the model achieves a median time to first audio of 88.9 ms. Serving is handled via SGLang-Omni, which requires Hopper-based GPUs for fused kernel validation.
The API is available at a list price of $50 per 1 million characters, with a promotional rate of $10 available until November 12, 2026. Users must specify the model header as drama-3 to avoid a fallback to the S2.1-pro version.
The model is released under the Fish Audio Community License v1.0. It is free for research and non-commercial use, as well as for commercial entities with annual revenue under $500,000. Third-party API hosting requires a separate license, and all products must include attribution.
