fishaudio/drama-3
main boardmodel
Fish Audio Drama 3 is Fish Audio's new-generation text-to-speech model, built for multi-character scenes and long-form content. Mixture-of-experts, 36B total, 3B active (256 experts, 8 per token). Context: 256K tokens (config). License: fish-audio-research-license.
- Org
- fishaudio
- Released
- 2026-10-01
- Status
- main board
- License
- fish-audio-research-license (restricted)
- Parameters
- 35.9B total, 3.0B active
- Architecture
- MoE · 256 experts, 8/token · FishOmniModel
- Context
- 262,144 tokens
- Task
- text-to-speech · audio
- Signal
- 1005 HF likes · 21375 downloads
Runs on
real quant files; need = weights + FP16 KV cache for 8K tokens + 1GB. A sizing calculation, not tested.
- Q4 ~25.5 GB: 32GB GPU (5090), Mac 36GB, Mac 48GB, CPU only, 32GB RAM (slow)
- Q8 ~40.2 GB: 48GB GPU (RTX 6000 Ada / A6000), Mac 64GB (M1 Max), Mac 96GB, CPU only, 64GB RAM (slow)
- FP16 ~73.3 GB: 96GB GPU (RTX Pro 6000 Blackwell), 1× DGX Spark (128GB), 2× DGX Spark (256GB), Strix Halo 128GB unified (~96GB usable), Mac 128GB, Mac 192GB, Mac 256GB, Mac 512GB, CPU only, 128GB RAM (slow), CPU only, 256GB RAM (slow)
Links
- Hugging Face
- quant: fishaudio/drama-3-int4
- MLX: mlx-community/fish-audio-drama-3-3bit
- MLX: mlx-community/fish-audio-drama-3-4bit
- MLX: mlx-community/fish-audio-drama-3-8bit
First published here 2026-10-01 · updated 2026-10-01 · markdown · JSON