Fish Speech releases Drama 3 — a 35B-parameter speech generation model with support for 80 languages

On October 13, 2026, a pre-release build of Fish Audio Drama 3 was published in the Fish Speech repository. It is a speech generation model that turns text into a "directed performance": it clones a voice from several references, follows natural-language instructions, voices scenes with multiple characters, edits speech via infilling, and synthesizes long or streaming audio. Details are in the release on GitHub. The Drama 3 model is built on a hybrid 35B-A3B MoE backbone with Gated DeltaNet linear attention: 256 experts, top-8 plus one shared expert, about 3B active parameters per token. The backbone is initialized from Qwen3.5. The audio decoder is a new codebook-MTP, 13 layers, 10 codebooks, a rate of about 21.5 Hz, output at 44.1 kHz via ModifiedDAC. Running it requires an NVIDIA GPU with at least 48 GB of VRAM in FP8 mode. The code is released as open source, and the model is pretrained on several hundred billion tokens across roughly 80 languages and dialects.

What it means

In our reference, Fish Speech is described as an open-source text-to-speech generator from Fish Audio. The current fact sheet refers to version 1.5: the fish-speech-1.5 model, the largest model weights file at 1.2 GB, all files at 1.4 GB. That is noticeably smaller than Drama 3, where the backbone alone has 35B MoE parameters and inference requires 48 GB of VRAM. The code license for Fish Speech is the Fish Audio Research License: free use for research and non-commercial purposes, with commercial use under a separate agreement with Fish Audio. The version 1.5 weights are distributed under cc-by-nc-sa-4.0. For Drama 3, the author does not specify a license in the release, so the terms for the new model are not yet defined. The fact sheet notes that S2 Pro natively supports SGLang acceleration: Continuous Batching, Paged KV Cache, CUDA Graph, and RadixAttention. The Drama 3 release states that S2-Pro remains fully supported with the same command-line interface, API server, and WebUI, and the model type is read from the checkpoint's config.json. This means the update does not break existing workflows for the previous version. For users who run Fish Speech on their own hardware, Drama 3 is a substantially heavier model: 48 GB of VRAM versus the previous 1.4 GB of model weights files. Those who do not have such a GPU are left with S2-Pro and its existing interface.