Fish Audio releases Drama 3: 35B MoE open-weight TTS, starting at 48GB VRAM
According to a MarkTechPost report dated October 14, 2026, Fish Audio released Drama 3, a mixture-of-experts text-to-speech model with 35B total parameters and about 3B activated per token, with weights open on Hugging Face under a community license capping commercial use at $500,000 in revenue, and FP8 self-hosting requiring more than 48GB of NVIDIA VRAM.
What changed
According to a MarkTechPost report dated October 14, 2026, Fish Audio released the Drama 3 text-to-speech model, saying the weights were published on Hugging Face alongside a Fish Audio hosted API. The report states the model has 35B total parameters with about 3B activated per token (8 of 256 experts plus 1 shared expert), with a backbone based on the Qwen3.5-35B-A3B language model, 40 layers grouped as 1 gated full-attention layer per 3 Gated DeltaNet layers, keeping only 1 layer of KV cache per 4 layers. The model supports performance instructions written in natural language or bracketed prompts, and the instruction language can differ from the spoken line's language; it supports 83 languages, and speaker tags can hold up to 16 characters in a single request, with the API accepting 1 voice ID per speaker. A new Codebook-MTP decoder replaces S2's frame-by-frame Fast AR decoder, feeding back only codebooks 0 and 1 to the backbone. The report says it scores 93.2% overall and ranks first on Artificial Analysis Pronunciation Robustness, versus 88.1% for the previous S2.1 Pro, with the weakest category being exact sequences such as code, paths, and emails at 86.3%. On hardware, the report says at least 48GB of NVIDIA VRAM is needed under FP8, with no support for CPU, Apple MPS, or AMD ROCm; official benchmarks used a single H200, with median time to first audio of 88.9 ms and real-time factor of 0.131 for a single stream, throughput of 927.6 frames/second across 16 streams, inference served via SGLang-Omni, and fused kernels validated only on Hopper GPUs. The license is the Fish Audio Community License v1.0: free for research and non-commercial use, free for commercial use under $500,000 annual revenue, separate licensing required for third-party API hosting, products must display Built with Fish Audio, and cloning real voices without consent is prohibited.
What it means for developers
For developers, the significance of Drama 3 lies in putting controllable performance instructions and multi-speaker script rendering into open weights; according to the source report, a single request can hold 16 characters and cover 83 languages, which opens the possibility that audiobook, dubbing, and scripted-dialogue pipelines could render an entire scene in one call. But open weights do not equal an open-source license: the source says the weights are offered under the Fish Audio Community License v1.0, commercial use is no longer free above $500,000 in annual revenue, third-party API hosting requires separate licensing, and products must display an attribution mark. This means packaging Drama 3 as an external voice service and using it internally are two different licensing paths, and you need to confirm your own revenue scale and distribution form first. The hardware threshold is also a constraint: the source says at least 48GB of NVIDIA VRAM is needed under FP8 with no support for CPU, MPS, or ROCm, and the official latency figures are based on a single H200, so self-hosting is not directly feasible for smaller teams that only have 24GB-class cards or non-NVIDIA environments. On quality, the source says its quality Elo (Drama 3 Preview 1242) exceeds the previous S2.1 Pro's 1190 but still trails Eleven v4 Turbo's 1328, indicating that instruction-following control and multiple speakers are its strengths while absolute audio quality is not the leading item in the source report. In addition, the source says the benchmark measures pronunciation rather than emotional expression, and emotional quality has only qualitative descriptions, so products whose main selling point is emotional performance should not base a judgment on that score alone.
Integration and evaluation
Recommendation one (model header and fallback protection, advisory analysis): the source reports that the API selects the model via the model header on POST /v1/tts, and a missing or misspelled header silently falls back to s2.1-pro. If you already have calling code for S2.1 Pro, the practical choice is to establish an explicit header constant with an assertion rather than rely on default behavior; this choice has the highest priority in existing pipelines that maintain multiple TTS versions at once, because a silent fallback distorts A/B results. A measurable evaluation criterion: on a regression set with fixed input text, verify that the model version field actually used in the response matches local records, with a fallback rate of 0, and compare whether pronunciation differences before and after the switch match the header. Recommendation two (the boundary between self-hosting and the hosted API, conditional analysis): the source says self-hosting requires more than 48GB of NVIDIA VRAM under FP8 and does not support CPU/MPS/ROCm, while the hosted API is priced at $50 per million characters, with the source saying a $10 launch price runs through November 12, 2026. The practical choice is to use the hosted API to absorb peak bursts first and evaluate self-hosting only when load is stable and use is compliant internal use; the premise is having a Hopper-class GPU on hand and annual revenue below the source's stated $500,000 threshold, otherwise the license and hardware constraints both apply at once. A measurable evaluation criterion: measure time to first audio and real-time factor at the target concurrency (for example the 16 streams mentioned by the source), calculate actual cost per million characters against self-hosting depreciation cost, and compare the crossover point on your monthly usage curve. Recommendation three (text preprocessing direction, advisory analysis): the source says the weakest category is exact sequences such as code, paths, and emails at 86.3%, while brands, places, and technical terms reach 95.9%. The practical choice is to rewrite or annotate the reading of code, path, and email fragments before sending them to TTS, rather than reading raw text directly; this choice pays off most in teaching and documentation narration content that includes technical terms and code examples. A measurable evaluation criterion: build a test set containing URLs, version numbers, and file paths, and compare character-level or human-listening error rates before and after preprocessing.
Limits and open questions
To be clear: this version, Drama 3, is not currently listed in KimiSeek's live model catalog; the API availability described in the source refers to Fish Audio's own hosted service, not this platform. The pricing in the source ($50 per million characters list price, $10 launch price through November 12, 2026) is report content from October 14, 2026; prices may have changed, so refer to Fish Audio's official pricing page and do not budget long-term on this basis. The source does not disclose context length; architecture and training details (the Codebook-MTP decoder, GRPO reinforcement learning, etc.) are relayed from the publisher and were not independently verified by this publication. On benchmarks, the pronunciation robustness score and quality Elo come from Artificial Analysis and Fish Audio self-reporting, and the Drama 3 Preview Elo is a preview-version result that does not represent the final version; emotional quality has only qualitative descriptions with no quantitative evidence. Hardware and latency data are based on a single H200 and fused kernels were validated only on Hopper, so actual performance on other GPUs is unknown. The final interpretation of the license terms depends on the original text of the Fish Audio Community License v1.0, and the exact scope of the revenue cap and third-party hosting requirements should be governed by the license text and legal advice. In addition, the 48GB VRAM threshold and the lack of CPU/MPS/ROCm support all come from the source and were not reproduced in a local environment.