Fish Audio open-sources 35.9B MoE TTS Drama 3, tops leaderboard while decoding faster than its 4.6B predecessor
kastnerkyle · x · 2026-10-19
Fish Audio released Drama 3 with open weights: a 35.9B-parameter MoE speech model (initialized from Qwen3.5-35B-A3B, 3B active in the backbone) that claims the largest total parameter count among open-weight speech models and takes 1st place on TTS leaderboards.
Architecture vs S2
- Hybrid Gated DeltaNet + MoE backbone replaces S2's dense Qwen3 model
- Only codes 0 and 1 of each frame feed back into the backbone (S2 summed all ten)
- New codebook-MTP audio decoder: a full-width causal-over-time layer per acoustic codebook, replacing the per-frame Fast AR decoder
- RL with a single learned human-preference reward — no ASR, speaker-similarity, or quality rewards
Performance: 7.8x the parameters of S2 (4.6B) yet faster — 271 vs 177 frames/s on one H200, thanks to weight reads dropping from 15.6 GB to 6.0 GB per frame and serial passes after the backbone cut from 40 to 13. Shipped with a technical report and launch-day SGLang-Omni serving. New tasks include output editing, scene dubbing, infill editing, and streaming text input.
More from Models
- Chris Potts critiques auto-interp evaluation: precision, recall, and the case for ex ante testing — ChrisGPotts · 2026-10-19
- Mistral releases Timestral time-series foundation models, up to 209M params — petrusenko_max · 2026-10-19
- Gemini 4 Scores Just 78% on Saturated FrontierMath Tier 4, Well Behind OpenAI's Models — haider1 · 2026-10-19
- OpenRouter Data Shows US Model Token Use Flat for 8 Weeks as Chinese Providers Hit 79.2T Weekly — SumitGup · 2026-10-19
- Anthropic reportedly tests its next model under Claude Opus 5.5, users spot big quality jumps — testingcatalog · 2026-10-19
- SWE-sweep: Meta's new benchmark shows top models fix under 5% of 4.1k real bugs — eliebakouch · 2026-10-19