Fish Audio releases Drama 3, a 35B-parameter MoE text-to-speech model with 16-speaker support

Fish Audio has released Drama 3, a text-to-speech model with 35B total parameters and 3B active per token, designed for multi-character scenes and long-form content. A single text can carry up to 16 distinct speakers. The weights are on Hugging Face under the Fish Audio Community License.

Architecture

According to the model card, Drama 3 pairs two components:

• Text backbone: a mixture-of-experts (MoE) model with hybrid attention, mixing full and linear layers. Total parameters are 35B, with 3B active per token. The Hugging Face metadata lists the checkpoint at 36B parameters.
• Audio decoder: a "Codebook-MTP" decoder that decodes 10 codebooks causally over the whole sequence.

MoE experts are stored in FP8 (F8_E4M3), and the remaining weights are in bf16.

Features

• Multi-speaker generation: speaker tags <|speaker:0|> through <|speaker:15|> let one text carry up to 16 characters.
• Voice cloning: the API takes a reference clip plus its transcript. The card's example uses a 3.9-second clip.
• Natural-language control and point tags: event tags such as [sighs] and [laughs] are written inline, at the position where the event should occur.
• Edit as continuation: a request can include previous_takes, the earlier text versions, so a revision such as changing "Wednesday" to "Thursday" or "noon" to "five" is generated in the context of prior takes. Edits can be chained.
• Long text: input is chunked automatically, and output can be streamed per chunk.

Serving

Drama 3 runs on SGLang-Omni from launch day. Fish Audio publishes a Drama 3 Cookbook with installation, deployment and streaming API examples.

Fish Audio's performance figures were measured on NVIDIA H200 GPUs with streaming requests on Seed-TTS English prompts. Each request cloned a voice from one pre-encoded reference and used temperature 1.0, top_p 0.9, top_k 30 and max_new_tokens 500. These are company-reported numbers, not independent results.

Concurrency (1 H200)	Throughput (frames/s)	TTFA p50	TTFA p95	Mean RTF
1	166.7	88.9 ms	90.8 ms	0.131
16	927.6	210.1 ms	287.2 ms	0.373
64	1,687.3	450.9 ms	1,020.1 ms	0.801

TTFA is time to first audio. RTF is request-to-last-chunk time divided by returned audio duration. Each cell is the mean of four runs.

Fish Audio also tested vocoder placement at higher concurrency. At 256 streams with the vocoder on the engine's GPU, throughput was 2,422 frames/s and median TTFA was 3,403 ms. Moving the vocoder to a second GPU raised throughput to 4,160 frames/s and cut median TTFA to 2,146 ms. Fish Audio describes this as about 72% more throughput and a 37% lower median TTFA. Each cell is the mean of six server starts.

Availability and pricing

• License: Fish Audio Community License. Check the license file for commercial terms.
• Pricing: not yet disclosed. No Inference Provider currently deploys the model.
• Context window, training data and quality benchmarks: not disclosed on the model page. Fish Audio links a technical report, which the page lists as published one day ago.
• Release date: not stated on the model page.

What this means

The notable design choice is a 3B-active MoE backbone on a TTS model. That keeps per-token compute low while the model carries 35B parameters of capacity, which probably explains the sub-100 ms single-stream first-audio latency. The 16-speaker tagging and edit-as-continuation interface target audiobook, drama and dialogue production, where regenerating a whole take for one changed word is costly.

The main gap is quality evidence. The model card publishes throughput and latency but no speaker-similarity, intelligibility or preference scores, so claims about naturalness cannot yet be compared with ElevenLabs or open TTS models. The deployment guidance is also notable: the second-GPU vocoder result suggests the vocoder, not the 3B-active backbone, becomes the bottleneck beyond roughly 128 concurrent streams. Teams planning high-concurrency serving should budget for that. The community license also means commercial use depends on terms that should be read before deployment.