Fish Audio Drama 3 Locally in ComfyUI: Run the 35B TTS on an 8 GB GPU

Fish Audio released Drama 3 on October 13, 2026, and it is the most interesting open weights text-to-speech model of the year so far: a 35B mixture-of-experts that can voice up to 16 characters in one script, takes stage directions like [voice drops to a terrified whisper], and tops Artificial Analysis’ pronunciation benchmark. The catch is that the official release is about 40 GB and is aimed at data-center GPUs. A day later, a community node pack called ComfyUI-Drama3 made it run on an 8 GB graphics card by keeping most of the model in system RAM. This guide explains what Drama 3 is, how much RAM and VRAM you really need, which file to download, how to install the ComfyUI nodes, how to write tagged multi-speaker scripts, and how to fix the problems people are hitting this week.

What Is Fish Audio Drama 3?

Drama 3 is Fish Audio’s new flagship speech model and the successor to its S2 line. Most open TTS models are built to read one voice at a time as cleanly as possible. Drama 3 is built for performances: radio plays, audiobooks with several characters, game dialogue and podcasts where people interrupt each other.

The key specs, from Fish Audio’s model card and launch coverage:

Architecturally, the text backbone mixes full and linear attention, and audio comes out of a codebook decoder that predicts 10 codebooks per frame. The practical upshot is that, like a modern MoE language model, only a small slice of the 35B parameters does work for each token. That is exactly what makes the 8 GB trick possible.

How good is it?

Fish Audio reports first place on Artificial Analysis’ Pronunciation Robustness benchmark, which uses human listeners rather than automatic transcription to score results.

Fish Audio also says it trained Drama 3 against a preference model instead of rewards based on how easy the audio is to transcribe. In plain terms, the model is not punished for sounding emotional, breathless or sarcastic, which older TTS training tends to flatten out.

Three Ways to Run Drama 3

Before you download 40 GB, decide which route you actually need.

The built-in Fish Audio nodes in ComfyUI core gained Drama 3 as a model option on launch day, next to s2.1-pro and s1. Those nodes need a Fish Audio API key and do not use your GPU at all. The rest of this guide is about the fully local route.

How Drama 3 Runs on an 8 GB GPU

A mixture-of-experts model has a shared core that every token passes through, plus many “expert” blocks of which only a few are picked for each token. In Drama 3 the shared, non-expert parts come to about 6 GB. The routed experts are the bulk of the 35B parameters.

ComfyUI-Drama3 keeps the 6 GB core on the GPU and leaves the experts in system RAM. When a token needs an expert, it is streamed into a cache in VRAM. Experts that are used often stay cached, so the bigger your cache, the less copying over PCIe and the faster it runs. This is the same idea as llama.cpp’s expert offloading for large MoE language models, applied to speech.

Two consequences follow:

• System RAM, not VRAM, is the real requirement. The whole model file has to fit in RAM with room to spare.
• More VRAM buys speed, not quality. An 8 GB card produces the same audio as a 24 GB card, just more slowly.

Hardware Requirements and Which File to Pick

The node author repacked Drama 3 into single files with different expert formats. Pick one, based mostly on your system RAM.

Why is INT8 faster than FP8 when the files are the same size? The FP8 experts get converted to bf16 on every pass, while INT8 has a direct kernel. On an RTX 3090 the author saw INT8 run about 14 to 17 percent faster.

The other requirements:

• GPU: NVIDIA only, 8 GB VRAM minimum. Measured peak VRAM ranged from about 6.7 GB to 21.8 GB depending on the expert cache size you choose.
• ComfyUI: version 0.40 or newer. The author tested on 0.41.0.
• Disk: about 26 GB for the W4A8 file plus the 1.6 GB codec, or about 40 GB for FP8 or INT8. Add 4 to 6 GB for the optional speech recognition models.
• A fast SSD helps. The first run reads the whole file from disk and can take minutes.

On a 16 GB card like my RTX 5080 with 64 GB of RAM, the INT8 file is the one I would pick: it fits in RAM with headroom, and a 16 GB card can hold a large expert cache.

Install ComfyUI-Drama3 Step by Step

1. Update ComfyUI to 0.40 or newer.
2. Clone the node pack into your custom nodes folder. No pip install is needed, because the CUDA kernels are compiled at runtime by PyTorch.

   cd ComfyUI/custom_nodes
   git clone https://github.com/PedroMarinhoDev/ComfyUI-Drama3

3. Download the model and codec. Swap the file name for the INT8 or W4A8 build if that suits your machine better:

   hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_int8experts_bf16.safetensors --local-dir ComfyUI/models/diffusion_models
   hf download PedroMarinhoDev/Drama-3-ComfyUI drama3_codec.safetensors --local-dir ComfyUI/models/vae

4. Optional: download speech recognition. Qwen3-ASR transcribes your reference clips so you do not have to type transcripts, and the forced aligner adds word timings to the editor:

   hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_asr_1.7b_bf16.safetensors --local-dir ComfyUI/models/asr
   hf download PedroMarinhoDev/Qwen3-ASR-ComfyUI qwen3_forced_aligner_0.6b_bf16.safetensors --local-dir ComfyUI/models/asr

   There is also a smaller qwen3_asr_0.6b_bf16 (1.6 GB) if disk space is tight.

5. Restart ComfyUI. The Drama 3 nodes appear in the node search.

If you are on Windows with the portable build, the paths are the same inside ComfyUI_windows_portable/ComfyUI/. Install the Hugging Face CLI with pip install -U huggingface_hub if the hf command is missing, or download the files from the repo page in your browser.

The Nodes and a Basic Workflow

A minimal workflow is: Load Drama 3 and Load Drama 3 Codec feed Drama 3 Generate, which feeds Drama 3 Decode, then a Save Audio or Preview Audio node. Add a Load Voice or Encode Voice node into a voice slot when you want a specific voice.

Cloning a Voice

Voice cloning works from a short reference clip and its transcript:

1. Load a clean clip with Load Audio. Cut it between sentences, avoid music and reverb, and pick a clip whose delivery matches the mood you want, since Drama 3 copies the energy of the reference.
2. Get the transcript. Type it, or run Qwen3-ASR Transcribe and correct anything it got wrong. The transcript must cover every word in the clip.
3. Run Drama 3 Encode Voice, then Save Drama 3 Voice. The saved voice is only a few kilobytes and does not need the original clip or the codec again.
4. Connect saved voices to the voice slots of Drama 3 Generate. Up to 16 speakers can share one script.

Only clone voices you own or have permission to use. The node author asks users not to clone an identifiable person’s voice without consent, and that is good practice anywhere.

Writing Scripts: Speakers, Tags and Directions

This is where Drama 3 earns its name. You direct it like an actor, with a small set of markup rules.

A short two-person scene looks like this:

<|speaker:0|>[calm, low voice] The lights went out at exactly midnight.
<|speaker:1|>[nervous laugh] That's... that's just the storm, right?
<|speaker:0|>Maybe. <emphasis>Or maybe</emphasis> someone wanted it dark.
<|speaker:1|>[voice drops to a terrified whisper] Did you hear that?

Directions do not have to be in the same language as the dialogue. Fish Audio’s own examples put Chinese, Spanish and Japanese directions on English lines. For long scripts, turn on split_long_text so the node breaks the text into chunks.

Two quirks to know: in “one take per speaker” mode, a line that opens with [long pause] can come out silent, and director’s notes that refer to an earlier take do not work, because the nodes do not send previous takes back into the prompt.

Speed: How Fast Is It Locally?

Drama 3 makes 21.5 audio frames per second of speech, so anything under 46 ms per frame is faster than real time. These are the node author’s measurements for a 9 second line in a cloned voice, dated October 17, 2026:

Fish Audio Drama 3 in ComfyUI: speed per audio frameLower is faster. Under 46 ms per frame is faster than real time.real time (46 ms)4090, W4A8, 16 GB cache11.7 ms3090+4090, experts in VRAM13.2 ms3090, INT8, 16 GB cache23.2 ms3090, FP8, 16 GB cache24.8 ms3090, INT8, 4 GB cache34.4 ms3090 at 8 GB, INT845.8 ms3090, FP8, 4 GB cache58 ms3090 at 8 GB, FP897.3 msSource: ComfyUI-Drama3 README, measured Oct 17 2026 on a 9 second cloned-voice line.

Three takeaways. First, an 8 GB card is right at real time with the INT8 file, which is genuinely usable for audiobook and video voice-over work. Second, the expert cache size matters more than the GPU model: the same RTX 3090 goes from 58 ms to 24.8 ms per frame with FP8 just by raising the cache from 4 GB to 16 GB. Third, the INT8 file beats FP8 most clearly when VRAM is tight. If you have two GPUs, putting the experts entirely in VRAM across both is the fastest option tested.

License and Responsible Use

There is some confusion here, so be careful. The Hugging Face metadata for fishaudio/drama-3 says “fish-audio-research-license”, while the model card text says the Fish Audio Community License. The ComfyUI-Drama3 README summarizes the weights and their outputs as covered by the Fish Audio Community License: free for research and non-commercial use and for commercial use by businesses under USD 500k annual revenue, with a separate license needed to run a hosted TTS service.

The node pack itself is GPL-3.0, with ported Fish Audio parts under Fish Audio’s license and parts adapted from Qwen’s ASR code under Apache 2.0. If you plan to monetize audio, read the actual LICENSE file in the Fish Audio repo rather than any summary, including this one.

Troubleshooting

The take goes silent after a second, or skips text

Set top_k to 0 (older workflows used 30), make sure the reference transcript covers every word in the clip, and use a clean clip cut between sentences. Trying a new seed often fixes a one-off silent take. A few reference clips fail even in Fish Audio’s own code, so try another clip if one keeps failing.

Voices swap between characters in long dialogue

An earlier chunk can pass its voice to the next one. Give every speaker in the script an explicit voice rather than leaving some on the default.

Speech sounds slow or flat

Keep temperature at 1.0 and top_p at 0.9 or higher. Lower values make the delivery monotone. Use a reference clip with the energy you want.

Out of memory, or the PC freezes while loading

If it is VRAM, lower expert_cache_gb. If it is system RAM, switch to the W6A8 or W4A8 file. Close browsers and other large apps during the first load.

The first run takes minutes

That is normal. The whole file is read from disk once. Later runs in the same session are much faster, and an NVMe SSD helps a lot.

FAQ

Can Fish Audio Drama 3 really run on 8 GB VRAM?

Yes, with the ComfyUI-Drama3 nodes and the experts in system RAM. The author tested an RTX 3090 limited to 8 GB and got about real-time speed with the INT8 file. You still need 32 to 48 GB of system RAM.

Does it work on AMD or Apple Silicon?

The node pack lists NVIDIA GPUs only. For other hardware, use Fish Audio’s API through the built-in ComfyUI partner nodes.

Can I run it with 16 GB of system RAM?

Not with any of the current files. The smallest, W4A8, is 24.6 GB on its own.

Which languages does Drama 3 support?

Fish Audio lists 83 languages. The bundled Qwen3-ASR transcriber supports 30, so for rarer languages you may need to type reference transcripts yourself.

How does it compare with Voicebox?

Voicebox is a lightweight desktop app that runs on almost any PC. Drama 3 is far heavier but handles multi-speaker scenes and stage directions that small models cannot. If your PC has under 32 GB of RAM, stick with a small model.

Can I use Drama 3 audio on YouTube?

According to the license summary in the node pack, commercial use is allowed for businesses under USD 500k annual revenue. Check the official license text and only use voices you have rights to.

Final Thoughts

Fish Audio Drama 3 is the first open TTS model that feels like directing a cast rather than feeding a text box, and ComfyUI-Drama3 makes it practical on ordinary gaming PCs. If you have an NVIDIA card with 8 GB or more and at least 48 GB of RAM, download the INT8 file and the codec, give it the biggest expert cache your VRAM allows, and start with a two-person scene. With 32 GB of RAM, the W4A8 file is your route in.

More local AI guides on this site:

• Voicebox Tutorial: Free Local AI Voice Generation on Windows, a lighter local voice option.
• Add Sound Effects and a Music Climax to a Short Without Re-encoding, to mix your Drama 3 dialogue into videos.
• Best Local AI Video Models in 2026, for visuals to go with your voice-over.
• Can You Run Meta Muse Spark Open Locally?, which uses the same expert offloading idea for a giant language model.

Sources

• GitHub: PedroMarinhoDev/ComfyUI-Drama3
• Hugging Face: fishaudio/drama-3
• ComfyUI Wiki: Fish Audio Drama 3 announcement
• ComfyUI Wiki: ComfyUI-Drama3 runs Drama 3 locally on 8 GB
• ComfyUI Docs: Fish Audio Drama 3 partner node