Evidence at a glance
Performance Becomes Part of the Input
Fish Audio released Drama 3 in October 2026 as a text-to-speech model for scripted dialogue, audiobooks, and dubbing. Creators can place bracketed directions before a line, such as “barely holding it together,” so the model generates not only the spoken words but also a prompted style of performance. Directions can be written in a different language from the dialogue, and paired tags can restrict an instruction to an enclosed span.
That change matters to technical leads because it turns some voice direction from repeated listening and parameter adjustment into a textual convention embedded in the source material. A script can identify up to 16 characters. The API takes one voice ID per speaker and can render a full scene in a single request. Drama 3 is therefore better understood as a speech-generation system with a performance-control interface, rather than just a sentence-to-audio endpoint.
A Large Model Does Not Mean Every Step Uses It All
Drama 3 has 35 billion parameters in total, but activates about 3 billion per token. The launch material says it selects 8 experts from a pool of 256, along with one shared expert. In this mixture-of-experts design, total model size and the number of parameters active at each step are different quantities. The 35B label alone does not tell you the compute cost of a generation request.
The backbone is based on Qwen3.5-35B-A3B and has 40 layers, arranged as three Gated DeltaNet layers followed by one full-attention layer. Only one quarter of the layers retain a KV cache, while the DeltaNet layers store history in a constant-size state. On the audio side, a new Codebook-MTP decoder uses four shared layers feeding nine layers, one per acoustic codebook. Only codebooks 0 and 1 feed information back into the backbone, compared with all 10 in S2. The supplied material does not include ablation results that would isolate this change’s effect on latency or audio quality.
Strong Pronunciation Scores Do Not Guarantee Every Read
Fish Audio reports that Drama 3 scored 93.2% overall on Artificial Analysis’s English Pronunciation Robustness test, ranking first and improving on S2.1 Pro’s 88.1%. It scored 95.9% on brands, places, and technical terms; 95.3% on context-dependent readings; and 93.6% on shorthand such as numbers, dates, and units. These results provide specific evidence about categories that are often mispronounced, rather than relying only on a subjective claim that the voice sounds more natural.
The same results show a boundary: exact sequences such as codes, paths, and email addresses are the weakest category, at 86.3%, though that is still above S2.1 Pro’s 79.4%. The benchmark measures pronunciation, not emotional delivery or character performance. Fish Audio also describes Drama 3 as its highest-scoring model on the cited Artificial Analysis measure, while noting that it remains behind several closed APIs on quality Elo. Teams should therefore evaluate performance direction and exact-string accuracy separately, rather than treating the former as a guarantee of the latter.
Open Weights Do Not Remove the Deployment Cost
Drama 3’s weights are available on Hugging Face, and the model can also be used through the Fish Audio API. Self-hosting has a substantial hardware floor: the supplied requirements call for an NVIDIA GPU with at least 48 GB of memory, with expert weights in FP8 and the remainder in bf16. CPU, Apple MPS, and AMD ROCm are not supported. Fish Audio reports results on one H200: 88.9 ms median time to first audio for one stream, an RTF of 0.131, and 927.6 frames per second at 16 streams. These figures are useful for sizing discussions, but should not be assumed to transfer to smaller GPUs or different serving configurations.
The hosted API has a migration detail that can otherwise go unnoticed: requests must set the model to drama-3 in the header, and omitting that header silently falls back to s2.1-pro. A request can therefore return audio without running the model the developer intended, so integration tests should verify model selection, not merely a successful response. The listed price is $50 per million characters, with a stated promotional price of $10 per million characters through November 12, 2026. Teams should budget against the standard price rather than treating the promotion as the durable cost.
Licensing Belongs in the Architecture Review
Open weights do not mean unrestricted use. Community License v1.0 allows research and non-commercial use, as well as commercial use by organizations with annual revenue below $500,000. Hosting the model as a third-party API requires a separate license. Users must also display “Built with Fish Audio,” and voice cloning without consent is prohibited. For teams embedding speech generation in a product that customers or end users will call, licensing and deployment choices should be reviewed together.
A practical evaluation can start with a representative script that includes speaker changes, emotional directions, brand or technical terms, and exact strings such as codes and email addresses. Teams can then record output quality, hardware requirements, and concurrency needs. The material does not disclose context length or provide an emotional-performance benchmark or comprehensive cross-hardware results, so it does not establish long-form continuity, complex emotional control, or low-cost deployment. If a workload fits single-request, multi-speaker script generation and the team can meet the hardware and license conditions, Drama 3 merits a controlled comparison. Open weights alone are not a sufficient reason to migrate.