Fish Audio opens Drama 3: speech model with 35 billion parameters, weights on Hugging Face
On October 13, 2026, the California company Fish Audio launched the Drama 3 text-to-speech model, with weights on Hugging Face and a promotional API price of $10 per million characters that expires on November 12. The model is meant to be steered with actor-style direction in plain language — but its emotional qualities are, so far, only the company's own description, and self-hosting requires a GPU with 48 GB of memory.
The News
Fish Audio, which according to its own press release builds tools for expressive text-to-speech, speech-to-text, and voice cloning, launched Drama 3 on October 13, 2026. The point is not just generating speech, but being able to direct it: the user describes tone, pace, and character in plain language, and the model is said to be able to shift emotion mid-sentence, among other things. The weights are on Hugging Face, and the model is in production in the company's own API.
In the launch announcement, co-founder and CEO Rissa Cao called the shift a step further: "Speech AI has become very good at generating speech. The next step is directing a performance," she said, according to the PR Newswire release carried by Yahoo Finance and Unite.AI. Co-founder and chief scientist Shijia Liao explained the technology this way: "Drama 3 uses a new architecture that delivers better control and better instruction following."
How the Model Is Built
According to the launch post, as reported by MarkTechPost, Drama 3 is a mixture-of-experts model with 35 billion total parameters, but roughly 3 billion active per token. For each token, 8 of 256 experts are activated, plus one shared expert — a choice that lowers the compute cost per generated second compared with a dense model of the same size.
The backbone starts from the Qwen3.5-35B-A3B language model and has 40 layers, grouped so that three Gated DeltaNet layers are followed by one gate-based full-attention layer. The decoder is also new: a Codebook-MTP decoder replaces the per-frame Fast AR decoder of its predecessor, S2. The context length is not disclosed in the technical reporting.
What Is Actually Measured — and What Is Merely Claimed
There are two types of performance figures here, and they say two different things. Fish Audio reports 93.2 percent overall on Artificial Analysis' Pronunciation Robustness benchmark, placing it first, versus 88.1 percent for predecessor S2.1 Pro. But that benchmark measures pronunciation robustness — not emotional expression or "acting," which is the model's main selling point. The emotional quality is only described qualitatively by the company itself.
On Artificial Analysis' quality leaderboard, "Drama 3 Preview" sits at an Elo of 1242 — above S2.1 Pro (1190), but below Eleven v4 Turbo (1328). The launch model and the "Preview" variant are not necessarily identical, so the number should be read as an indicator, not a final verdict.
Another unresolved matter is audio inpainting — the ability to mark a segment of a performance, change the instructions, and regenerate only the selected part while the rest of the audio is preserved. The press release and MarkTechPost describe inpainting as a launch feature, while Notebookcheck writes that both Interactive Editing and Audio Inpainting are still in limited, invitation-based testing. In other words, the status is unclear in the source material. In any case, Unite.AI points out that the release contains no independent evaluation of how reliably single-word repair preserves surrounding audio across voices, accents, and difficult transitions.
What It Costs to Use — and to Self-Host
The list price in the API is $50 per million characters, with a promotional price of $10 in effect through November 12, 2026. The weights use the Fish Audio Community License v1.0: free for research and commercial use as long as annual revenue is under US$500,000, with attribution and consent requirements.
Self-hosting has a clear floor: the installation guide requires at least 48 GB of NVIDIA GPU memory in FP8, and CPU, Apple MPS, and AMD ROCm are not supported. The official benchmarks were run on a single H200. That places self-hosting in the price class of server GPUs, not workstations.
According to Notebookcheck, Drama 3 supports multi-speaker conversations and 83 languages, and is available via the website, the mobile apps, and the API.
Open Questions
Three things remain to be clarified. First, all technical figures and performance claims trace back to Fish Audio itself — via a paid press release and secondary reporting that cites the same launch post — with no independent verification. Second, the flagship features, such as mid-sentence emotion shifts and single-word inpainting, lack independent measurement. Third, growth figures in the press release, such as ARR and user counts, are unverified marketing claims and should be treated as such.
What It Means Before November 12
For developers, the choice is primarily one of timing. At $10 per million characters until November 12, the discounted API is the cheapest way in, and it offers a chance to test the contested emotional qualities on your own material before committing. Self-hosting gives full control and no per-character cost, but requires a 48 GB NVIDIA GPU and acceptance of the license's revenue cap and consent requirements — a choice that only pays off at consistently high volume or with strict requirements around data and infrastructure. The hard numbers available so far mostly concern pronunciation; whether the "direction" works as promised, experience with the model will have to decide.