Drama 3

ArtRealmAI reports the full October 13, 2026 launch of Fish Audio Drama 3 with open weights on Hugging Face and SGLang-Omni support from day one, describing the model as "open-weight rather than open source" and noting a $52M seed funding round announced alongside the launch. The article documents in-line square-bracket direction that alters the read mid-sentence, paired word tags such as whisper that scope a direction, 83 supported languages, and up to 16 speakers per script via speaker number tags. The same piece cites a 93.2% score on Artificial Analysis' Pronunciation Robustness benchmark versus 73.2% for S2-Pro, with the largest gains on numbers, units, file paths, and emails. Under the hood it describes a Qwen3.5-35B-A3B-derived backbone with hybrid attention arranged as three Gated DeltaNet layers per one full-attention layer, a codebook-MTP decoder over S2's 44.1 kHz DAC tokenizer, and GRPO preference tuning without a KL penalty.

ComfyUI Wiki confirms Drama 3 as Fish Audio's 35B MoE text-to-speech model announced on October 13, 2026, with 3B parameters active per token, hybrid full plus linear attention, FP8 expert weights, and a 10-codebook causal codebook-MTP decoder distributed across eight safetensors shards plus a 1.9 GB codec, roughly 40 GB on disk. The article documents speaker tags for up to 16 characters, automatic long-form chunking with per-chunk streaming aimed at audiobooks and dubbing, and support for 83 languages. The same page describes square-bracket direction written in plain language that can be placed anywhere in the line so the read changes where the mood shifts, with cross-language direction demonstrated using Chinese, Spanish, and Japanese cues on English text. It reports Fish Audio's claim of first place on Artificial Analysis' Pronunciation Robustness benchmark and notes that Comfy-Org/ComfyUI shipped a Drama 3 option in the Fish Audio partner node on the same day as the launch.

Fish Audio's own speech documentation site lists Drama 3 as the new-generation 35B-A3B flagship with 3B parameters active per token, alongside the previous S2-Pro 4B release, framing it as an upgrade across long-form audio, interaction, and editability. The page advertises multi-turn director revision, where the previous take can be guided with phrases like "gentler, and a bit sadder" and is regenerated in context rather than from scratch. Infilling and precise word-level editing, such as swapping "Wednesday" for "Thursday" while preserving surrounding timbre, are highlighted as headline features. The same documentation connects Drama 3 to the Fish Audio Community License and routes users to SGLang-Omni for efficient serving, with deployment paths for command line, API server, WebUI, and Docker. It positions the model as the successor to S2-Pro within Fish Speech while keeping the same command-line and API surface, and links to the Hugging Face weights plus the Drama 3 Technical Report PDF for deeper specification detail.

Drama 3 is Fish Audio's new-generation open-weight text-to-speech model built around a 35B-parameter mixture-of-experts text backbone with about 3B parameters active per token, hybrid full plus linear attention, and a codebook-MTP decoder that produces 10 codebooks causally. The Hugging Face listing reports FP8 MoE expert weights with the remaining tensors in bf16, multi-speaker generation through speaker 0-15 tags for up to 16 characters in one stream, and automatic chunked streaming for long-form text. The model is live on SGLang-Omni from launch day. The Hugging Face model card documents production serving numbers measured on NVIDIA H200 GPUs with streaming Seed-TTS English prompts, showing a single stream's time-to-first-audio at about 89 ms and throughput rising from 166.7 to 1687.3 frames per second across concurrency 1 to 64. Speaker tags, natural-language direction, and codebook-level continuation editing are demonstrated in working code samples on the same page. The model's release assets also ship under the FISH AUDIO COMMUNITY LICENSE.