Drama 3 Debuts: A 35B MoE Speech Model Tuned for Directable Acting
Fish Audio has officially released Drama 3, an open-weight directable speech synthesis model designed to move voice generation beyond flat line reads into fully expressive acting. Built on a Qwen3.5-35B-A3B Mixture-of-Experts (MoE) backbone, the system houses 35 billion total parameters while activating approximately 3 billion parameters per token.
Architecturally, the model utilizes 40 layers structured in an asymmetric layout: three Gated DeltaNet layers for every single gated full-attention layer. Because only one in four layers maintains a key-value (KV) cache, memory scaling during extended sequence decoding remains restrained. The architecture pairs this layout with a causal Codebook-MTP decoder for streaming generation.
The flagship capability of Drama 3 is natural-language stage direction. Developers can insert emotive point cues such as [sighs] alongside paired word-level delivery tags like <whisper> to dictate cadence, intensity, and physical vocalizations. The model also handles multi-speaker dialogue within a single generation request, accommodating up to 16 distinct voices tagged from <|speaker:0|> to <|speaker:15|>.
Hardware Footprint, Accuracy Benchmarks, and Licensing Framework
Deploying Drama 3 locally requires high-memory infrastructure. In FP8 precision, the model demands an NVIDIA GPU setup with at least 48 GB of VRAM. An alternative quantized int4 checkpoint reduces the raw weights footprint from 37.0 GB down to 20.8 GB, offering operational flexibility for enterprise teams running tighter hardware configurations.
On benchmark evaluations, Drama 3 scored 93.2% overall on the Artificial Analysis Pronunciation Robustness benchmark, achieving 95.9% accuracy on technical jargon and brand nomenclature, alongside 86.3% on exact-sequence tracking. In comparative evaluations, the model registered a quality Elo score of 1242, backed by zero-shot and reference voice cloning across 83 languages.
Fish Audio published the model under the Fish Audio Community License v1.0. This license grants free usage for non-commercial research as well as commercial enterprises with gross annual revenues below $500,000. For organizations choosing fully managed cloud access, Fish Audio lists its hosted API at $50 per 1 million characters, temporarily discounted to $10 per 1 million characters through November 12, 2026.
Developer and Practitioner Reactions: Praising Stage Direction, Debating Hardware Limits
Early practitioner reaction has highlighted Drama 3's dramatic potential, particularly among creators testing automated multi-character storytelling. Developers praised the ability to generate complete dialogue exchanges between up to 16 individual personas within a single API query, noting that native acting cues eliminate substantial post-production splicing and audio engineering overhead.
At the same time, community sentiment flagged clear operational trade-offs. The 48 GB VRAM threshold for running unquantized FP8 weights presents a barrier for independent practitioners and small studios lacking enterprise workstation infrastructure. Reviewers also noted that while a 1242 quality Elo score is competitive for an open-weight model, it still trails the perceived fidelity ceiling established by proprietary closed-source speech providers such as ElevenLabs.
Technical discussions have also centered on platform portability. At launch, production serving relies strictly on NVIDIA CUDA through SGLang-Omni pipelines. Native execution across alternative environments—such as AMD ROCm, Apple MPS, or CPU inference—remains unconfirmed and unsupported in the initial repository documentation.
Strategic Implications for Thai Media, Gaming, and Enterprise Operations
For Thai media enterprises, interactive gaming studios, and localization firms, Drama 3 represents an inflection point in automated voice production. Rather than paying recurring per-character fees to foreign closed-API providers, creative teams can explore local orchestration of complex audio dramas, marketing materials, and localized adaptations across the model's 83 supported languages.
The licensing terms offer meaningful headroom for the local tech ecosystem. Thai startups and digital agencies generating under $500,000 annually can deploy Drama 3 into customer-facing applications—ranging from conversational interactive voice response (IVR) to dynamic video game NPCs—without licensing royalties, preserving capital during early commercial scaling.
Nevertheless, technology leaders in Thailand must weigh infrastructure expenses before committing to on-premise deployments. Provisioning enterprise GPU capacity capable of meeting the 48 GB VRAM requirement incurs substantial cloud hosting costs. Organizations should also independently audit Thai tonal inflections, contextual pacing, and local brand phrasing against their internal brand standards before retiring existing human-assisted pipelines.
Enterprises in media, gaming, and customer operations gain a self-hostable voice engine that can act out entire dramatized scripts with fine-grained performance cues under an accessible community license.