Running Fish Audio Drama 3 on a 16GB AMD GPU: Red Curtain ports the 35B MoE that officially wants 48GB
fechyyy · reddit · 2026-10-18
Drama 3 officially demands a 48GB NVIDIA GPU, but Redditor fechyyy built Red Curtain (with Claude's help) to run the 35B MoE (3B active per token) on a 16GB RX 9070.
- FP8 experts (32GB) stay in system RAM; only used experts get cached on GPU, rest runs BF16 on-card
- 17s load, 11-12GB VRAM, 35GB RAM; 3.3s of speech in 13s, 11.6s in 37s — slower than realtime but fine for short takes
- Web UI with voice cloning from a reference clip, multi-speaker scripts, and Fish's editor tags ([laughing], <whisper>, etc.)
- The "first line then silence" bug is a known Drama 3 issue, not the AMD port; fixes: top-k off, per-speaker reference voices, no sound tags at line starts
- Tested only on RX 9070 (gfx1201); weights download from Hugging Face
More from Infra
- TSMC: 'No intention to leave any food on the table for others' — Beth_Kindig · 2026-10-18
- DeepInfra cuts DeepSeek V4.1 Flash price 51% as agent demand surges — gharik · 2026-10-18
- a16z avoids neoclouds over debt risk; Bittensor pitched as the answer to CoreWeave's model — markjeffrey · 2026-10-18
- SemiAnalysis: vLLM hits 1.3B tokens per $ TCO, 4x better than B300 aggregated serving — woosuk_k · 2026-10-18
- Local AI users rant: 16GB VRAM GPUs can't run any decent modern models anymore — 2MuchNonsenseHere · 2026-10-18
- Meta lays out 448G in-rack interconnect roadmap: PAM6 candidate, active copper cables stretch reach — StewartalsopIII · 2026-10-18