AEmotionStudio/drama-3-models

Fish Audio Drama 3 (int4) — mirror for Mæstræa

Built with Fish Audio.

A mirror of Fish Audio's Drama 3 — the official int4 checkpoint, fishaudio/drama-3-int4 @ a6f58869 — laid out for the Mæstræa AI Workstation (AI ▸ Voice ▸ Drama 3). All credit for the model goes to Fish Audio: see the Drama 3 model card and fish.audio/drama-3.

License

These Fish Audio Materials are licensed under the Fish Audio Community License, Version 1.0. Copyright © 39 AI, INC. All Rights Reserved. The full agreement is in LICENSE; NOTICE carries the required attribution and lists what changed. Nothing here is relicensed.

In short — read LICENSE for the binding terms:

• Free for research and non-commercial use.
• Commercial use is free while you and your affiliates earn under USD 500,000 a year; above that you need a licence from Fish Audio (business@fish.audio).
• Offering Drama 3 as a hosted speech service (an API, or an app whose main purpose is TTS, voice-over or cloning for others) always needs a separate licence.
• Never clone or imitate a real person's voice without their consent, never use it to impersonate, defraud or harass, and never remove watermarks or provenance signals. Anyone you make Drama 3 speech generation available to must be bound to the same restrictions.
• Don't use the outputs to train other speech, audio or foundation models.
• Products built with it must show "Built with Fish Audio".

By downloading these files you accept the Fish Audio Community License.

What is here

File	What it is
model-000NN-of-00012.safetensors, model.safetensors.index.json	The int4 checkpoint (routed experts int4 GPTQ-symmetric, everything else bf16) — byte-identical to upstream
codec.safetensors	The 44.1 kHz audio codec — upstream's codec.pth converted to safetensors, every tensor bit-identical
config.json, tokenizer*.json, special_tokens_map.json, chat_template.jinja	Byte-identical to upstream
LICENSE, NOTICE, sha256.json	The agreement (verbatim), the attribution + changes, and the sha256 of every file above

Fish's evaluation folders, samples and quantization scripts are not included — use the upstream repository for those.

How Mæstræa runs it

Mæstræa loads these weights with its own Apache-2.0 inference code (backend/ai/drama3_vendor, ported from sglang-omni and vllm-omni); no fish-speech code is used. On a 12 GB GPU the dense layers, codebook tower and codec decoder stay resident while the 16.6 GB of int4 experts stream from the memory-mapped shards through a GPU cache — about 2–3× real time on an RTX 4070. Scenes render line by line so each character keeps their own voice, and the main token is sampled with top-k off (a top-k cap traps cloned voices in silence).