--- title: "Can You Run Fish Audio Drama 3 Locally? VRAM and Setup Guide" url: "https://goremy.ai/blog/articles/fish-audio-drama-3-run-locally-vram" author: Luis Chavez-Mattos editor: Luis Chavez-Mattos editor_title: Director, Product published: "2026-10-23T00:20:06.296Z" updated: "2026-10-23T00:20:06.296Z" topic: Local & Open-Weight Models description: Fish Audio Drama 3 is an open-weight MoE text-to-speech model. Here are the VRAM needs, the 4-bit quant, bracket directions, and how it ranks. --- # Can You Run Fish Audio Drama 3 Locally? VRAM and Setup Guide > Fish Audio Drama 3 is an open-weight MoE text-to-speech model. Here are the VRAM needs, the 4-bit quant, bracket directions, and how it ranks. Edited by Luis Chavez-Mattos, Director, Product ยท Published 2026-10-23 ## Can you run Fish Audio Drama 3 locally? Yes, if you have a workstation-class GPU. Drama 3 is an open-weight text-to-speech model on Hugging Face. Community reports put the memory need at roughly 46 GB of VRAM, or about 32 GB with a community 4-bit quantized build. Fish Audio's model card lists about 35 billion parameters, with 3 billion active per token. Most laptops and mainstream gaming cards will not fit it. A 48 GB card or a 32 GB card with the quant is the realistic floor. ## TL;DR - Drama 3 is an **open-weight text-to-speech model** from Fish Audio, built for multi-character scenes and long-form audio. - The model card lists **35B parameters with 3B active per token** (a mixture of experts design), while a reviewer's walkthrough called it 36B. - Reported memory needs are **about 46 GB of VRAM**, or around 32 GB with a community 4-bit quant. Fish Audio's card does not state a VRAM figure. - You direct delivery with **plain-English notes in square brackets**, such as `[sighs]` or a short acting instruction, instead of picking an emotion preset. - One script can carry **up to 16 speakers** using speaker tags. - Results are mixed: it **tops Artificial Analysis on pronunciation** but did less well in blind arena preference voting. ## What is Fish Audio Drama 3? Fish Audio describes Drama 3 as its new-generation text-to-speech model for multi-character scenes and long-form content. It supports multi-speaker generation, voice cloning, and natural-language control. The architecture details from the model card: - A mixture-of-experts text backbone with hybrid (full plus linear) attention. - A codebook-MTP audio decoder that decodes 10 codebooks causally across the sequence. - FP8 MoE experts, with the remaining weights in bf16. - Automatic chunking of long text, which can be streamed chunk by chunk. The weights are on Hugging Face under the Fish Audio Community License. A reviewer summarized the terms as free for commercial use below roughly half a million dollars in annual revenue. Read the license file yourself before shipping anything, because the model card only points to it. ## How much VRAM does Drama 3 need? The model card gives no VRAM number. The figures that circulate come from community reports: | Setup | Reported VRAM | |---|---| | Standard weights | Roughly 46 GB | | Community 4-bit quantization | About 32 GB | Treat both as approximate. They are second-hand and will shift with context length, batch size, and the serving stack. The "3B active" detail can mislead. In a mixture-of-experts model, only a few experts run per token, which cuts compute. The full set of weights still generally has to sit in memory, so the memory footprint tracks the total parameter count, not the active count. That fits the reported numbers: a 35B model needs a big card even though each step is cheap. In practice, a 48 GB card (an RTX 6000 Ada class workstation GPU, for example) covers the standard weights. A 32 GB card covers the quantized version. Quantizing a speech model can change audio quality in ways text benchmarks do not capture, so listen to the output before committing to it. ## How do you run Drama 3 locally? The model card documents two serving paths. Both expose an API server you then call with text and optional reference audio. SGLang-Omni: ```bash git clone https://github.com/sgl-project/sglang-omni.git && cd sglang-omni uv pip install --prerelease=allow -v -e ".[fishaudio]" sgl-omni serve --model-path fishaudio/drama-3 --port 8000 ``` vLLM-Omni: ```bash git clone https://github.com/vllm-project/vllm-omni.git && cd vllm-omni uv pip install -e . vllm serve fishaudio/drama-3 --omni --port 8091 ``` Both projects publish a Drama 3 cookbook or recipe with deployment and streaming examples. The community 4-bit build is separate from these official instructions, so follow whatever instructions come with that release. ## How does the bracket direction work? Most TTS tools give you a dropdown: happy, sad, angry. Drama 3 lets you write the direction the way you would brief an actor, in square brackets inside the script. A few patterns from the model card and demos: - **Point tags.** Event tags go where the sound should land. The card's example is `[sighs] Fine. If we miss it, it's on you. [laughs] Well, maybe that's fair.` - **Delivery notes.** A demo showed a line with a note like "barely holding it together, forcing a smile" ahead of the spoken text. Same words, different performance. - **Word-level direction.** The demo also whispered one phrase inside an otherwise normal line. - **Sound effects.** Small laughs and shaky breaths can be placed exactly where you want them. Fish Audio says takes vary from run to run. Plan to generate a few and keep the best. ## Does Drama 3 support multiple speakers and editing? Yes to both. Speaker tags run from `<|speaker:0|>` to `<|speaker:15|>`, so one text can carry up to 16 characters. That is the main reason it targets audiobooks, game dialogue, and scripted scenes. Fish Audio licensed Uncle Roger's voice for its demo page, which shows the multi-voice and cloning features in action. The card also describes "editing as continuation." You send the earlier take (as `previous_takes`) along with a revised text, and the model regenerates with that context. The example changes "Wednesday" to "Thursday" in a sentence, then chains a second edit changing "by noon" to "by five." It works like prompting a language model with prior output. You get small word changes without redoing the whole line. Cloning needs a short reference clip plus its transcript. The card's example uses a 3.9 second clip. ## How fast is Drama 3? Fish Audio's numbers come from NVIDIA H200 GPUs, not consumer hardware, so they show the ceiling rather than what you will see at home. On one H200 with streaming requests: | Concurrency | TTFA p50 | RTF (mean) | |---:|---:|---:| | 1 | 88.9 ms | 0.131 | | 16 | 210.1 ms | 0.373 | | 64 | 450.9 ms | 0.801 | TTFA is time to first audio. RTF is generation time divided by audio duration, so anything under 1.0 is faster than real time. A single stream starts speaking in under 100 ms. The card also reports that moving the vocoder to a second GPU raised throughput about 72% at 256 streams. That matters for servers, not for a single local user. ## How does Drama 3 score against other TTS models? The results are mixed. On Artificial Analysis, Drama 3 scored highest on pronunciation, ahead of ElevenLabs and SpaceX AI's text-to-speech model, according to a reviewer's check of the leaderboard. In the arena leaderboard, where listeners pick favorites in blind comparisons, it did not rank as well. The practical reading: it handles hard words, names, and tricky text very well, but many listeners still prefer the sound of other voices. If your content is full of unusual terminology, pronunciation may matter more. If you want the most pleasing default voice, test it against alternatives with your own scripts. ## What does the hosted API cost? If local hardware is out of reach, Fish Audio runs a hosted API. A reviewer reported about $10 per million characters at launch, rising to $50 after November 12th. Verify current pricing on Fish Audio's site. One warning from the same review: the API can quietly fall back to older models and still bill you. Explicitly set the model to Drama 3 and confirm it in your requests. ## Frequently Asked Questions ### What GPU do I need for Fish Audio Drama 3? Community reports say roughly 46 GB of VRAM for the standard weights. A community 4-bit quantized version reportedly fits in about 32 GB. Fish Audio's model card does not list a VRAM requirement, so treat these as estimates. ### Is Drama 3 36B or 35B parameters? The Hugging Face model card says 35B parameters with 3B active per token. A reviewer described it as 36B. The model card is the primary source, so use 35B. ### Can I use Drama 3 commercially? It ships under the Fish Audio Community License. One reviewer summarized it as free for commercial use below roughly $500,000 in annual revenue. Check the license file for the exact terms before building a product on it. ### Does Drama 3 support voice cloning? Yes. The model card lists voice cloning, and its examples clone a voice from a short reference clip with a matching transcript. It also supports up to 16 speakers in one text through speaker tags.