Fish Audio Releases Drama 3: A 35B Open-Weights MoE Text-to-Speech Model

The speech synthesis space has gained another open-source option worth watching. Fish Audio has officially released Drama 3 — a Mixture-of-Experts (MoE) text-to-speech model with 35 billion total parameters and roughly 3 billion active parameters. This means it delivers large-model-level speech generation at inference costs far below dense models of comparable scale, and its weights are fully open.

Core Specs of Drama 3

The most striking thing about Drama 3 is its architectural choice. 35B total parameters might sound like it would consume an entire H100, but the MoE design means only about 3B parameters are activated per inference. This sparse activation strategy has already proven effective in the LLM world, and Fish Audio is bringing it to the TTS domain.

For developers, the practical implication is straightforward: you get a model with a knowledge capacity approaching 35B, but inference costs on par with 3B. In speech applications that demand low latency and high concurrency, this combination is more practically valuable than simply stacking up parameter counts.

The model covers 83 languages — a number that puts it in the top tier of open-source TTS models. Multilingual capability is no longer just "supporting" a few major languages; it's genuinely aimed at global deployment scenarios.

What is MoE (Mixture of Experts)?

MoE (Mixture of Experts) is a neural network architecture whose core idea is to split a large feed-forward network into multiple parallel "expert" sub-networks, then use a "router" (router/gating network) to dynamically select and activate only a few experts based on the input. This way, the model's total parameter count can be very large (high storage capacity), while the parameters actually involved in computation per forward pass are few (low compute).

In traditional dense models, total parameters equal compute. MoE decouples "total parameters" from "active parameters." Switch Transformer, Mixtral 8x7B, and similar models are classic examples: large total parameter counts, but only a subset of experts activated each time. This is exactly how Drama 3 achieves 35B total parameters with ~3B active — and the key to its inference costs being on par with small models. The trade-off is that VRAM must still hold all expert weights, and router training faces challenges like load balancing.

Natural Language Instructions: Control Speech by Speaking

Another design orientation of Drama 3 is the simplification of control. The model accepts "plain-language direction" rather than requiring professional parameter tuning like voice tags or parameter panels.

This approach aligns with the direction of speech generation research over the past two years — shifting the control interface from structured parameters to natural language. Users can describe the desired tone, emotion, and rhythm in everyday expressions, and the model translates those descriptions into acoustic features. While lowering the barrier to use, it also shifts the hard problem of "how to describe voice" from the user to the model side.

How well this works in practice depends on the model's semantic understanding of instructions. The MoE architecture may help here: different experts can handle different speech styles and language features, making it easier to decompose and route the multi-dimensional information in instructions.

Open Weights and Revenue-Capped License

Drama 3 is released under open weights with a "revenue-capped community license."

This is a type of compromise for commercializing open-source models. Open weights mean researchers and developers can download, fine-tune, and deploy, but commercial use has a revenue threshold: once revenue exceeds a certain scale, a separate license is required. Compared to a pure research license, it gives startup teams and small products ample room; compared to a fully permissive license, it preserves a path for the model creators to profit from large-scale commercial use.

For startup teams, this licensing model is worth reading carefully — especially how "revenue" is defined, and whether it's calculated based on the company's overall revenue or the revenue of the specific product line.

The Difference Between "Open Weights" and "Open Source"

Strictly speaking, "open weights" is not the same as "open source." Open weights means the trained parameter files are publicly downloadable — anyone can load, infer, and fine-tune. But training code, training data, and some license terms may not be fully open, differing from OSI-defined open-source licenses.

In recent years, many large models (like the Llama series) have adopted this model: weights are accessible but subject to additional terms such as user count limits, commercial restrictions, naming requirements, or revenue thresholds. That's why the industry often calls these "open-weight models" rather than "open-source models." Understanding this distinction matters for developers — it directly affects whether you can compliantly integrate the model into commercial products, and whether you need to sign a separate agreement with the model provider.

What This Release Means

Placing Drama 3 in the current open-source TTS landscape, several signals stand out.

First, MoE is spreading from language models to speech models. Sparse activation isn't exclusive to LLMs — any generative task that needs large knowledge capacity but is constrained by inference budgets could benefit from this architecture.

Second, open-weight TTS models at the 35B level remain rare. Most open-source speech models stay in the range of a few hundred million to a dozen billion parameters, and Drama 3 pushes the ceiling of open weights up a notch.

Third, the combination of multilingual coverage and natural language control points to a positioning as a "universal speech interface" rather than a single-scenario dubbing tool.

A caveat: model capability ultimately needs to be validated through actual listening experience and stability. Parameter scale, language count, and license terms are all on-paper metrics. The real bar is prosodic consistency in long-form text, naturalness of cross-language switching, and precision of instruction control. These still need to be answered by the community through real-world use.

For teams currently evaluating speech synthesis solutions, Drama 3 offers a new candidate: if you need multilingual support, flexible style control, and want controllable, fine-tunable weights, it deserves a spot on your testing list.

Background Notes

Common Design of Revenue-Capped Licenses

"Revenue-capped community licenses" are common in AI model releases, typically stipulating: individuals and teams with annual revenue or funding below a certain threshold (e.g., $1 million) can use it commercially for free; those exceeding the threshold need to apply for a commercial license. When reviewing such terms, focus on three things: first, the definition of "revenue" — company-wide revenue or product-line revenue for the model; second, whether the threshold is calculated annually, monthly, or cumulatively; third, whether distribution and commercial use of derivative models (fine-tuned weights) are subject to the same constraints. These details often determine whether small and medium teams can use it long-term at low cost.

One-line Summary

Fish Audio open-sources Drama 3, a 35B-parameter MoE text-to-speech model with only ~3B active parameters.

Paragraph Summary

Fish Audio released the open-source text-to-speech model Drama 3, using a Mixture-of-Experts (MoE) architecture with 35 billion total parameters and about 3 billion active. It delivers large-model-level speech generation at inference costs close to small models. The model covers 83 languages, supports natural language instructions to control tone, emotion, and rhythm, and is released under open weights with a "revenue-capped community license." The article argues this signals MoE spreading from language models to speech models, pushing the parameter ceiling of open-weight TTS higher and pointing toward a "universal speech interface" positioning. However, it also cautions that parameters, language counts, and license terms are only on-paper metrics — long-form prosodic consistency, cross-language naturalness, and instruction control precision still need community validation.