What We Mean by Open Source, and Why It Matters for S2
Mar 12, 2026
Rissa Cao, CEO
Since we released S2, the most common question hasn't been about benchmarks or architecture. It's been about the license.
"Can you clarify what you mean by 'open source'? Because I can see it's not for commercial use."
Fair question. Here's our answer.
What We Released
With S2, we released the full set of components needed to run, study, and build on the model:
- Model weights: the full 4B-parameter Dual-AR model
- Fine-tuning code: train on your own data, on your own infrastructure
- Production inference engine: via SGLang-Omni, the same stack we run in production
- Full technical report: architecture details, training recipe, benchmark methodology Download it. Run it locally. Fine-tune it. Inspect every layer. It's all there.
What the License Says
S2 is released under the Fish Audio Research License.
- Research and non-commercial use: fully free. No restrictions.
- Commercial use: requires a separate license from Fish Audio. No hidden clauses, no retroactive restrictions.
Open Source vs. Open Weights: Where S2 Sits
We want to be straightforward about this: S2 is open weights, not open source by the OSI definition.
In the AI industry today, the term "open source" covers a wide spectrum of release models. Every organization makes different trade-offs to balance community access with business sustainability. We chose our current license model to ensure we can continue funding our R&D while still sharing valuable tools.
Instead of debating labels, we want to give you full transparency on exactly what we provide. To help clarify our approach, here is a breakdown of how the S2 release compares to other major models in the space:
| Released Components | S2 | Llama 4 | DeepSeek R1 | Mistral Large 3 | GPT-OSS |
|---|---|---|---|---|---|
| Model weights | ✅ | ✅ | ✅ | ✅ | ✅ |
| Fine-tuning code | ✅ | ✅ | ✅ | ✅ | ❌ |
| Inference engine | ✅ | ❌ | ❌ | ❌ | ❌ |
| Technical report | ✅ | ✅ | ✅ | ✅ | ❌ |
| Free commercial use | ❌ | ✅ (< 700M MAU) | ✅ (MIT) | ✅ (Apache 2.0) | ✅ |
| Training data | ❌ | ❌ | ❌ | ❌ | ❌ |
We believe this is one of the more complete releases in the TTS space. Beyond weights and a paper, we also released fine-tuning code and the production inference engine, which is uncommon at any scale.
Why We Chose This License
Building and maintaining a state-of-the-art TTS model requires sustained investment in training, data infrastructure, and research. As a startup competing in a market with some of the largest technology companies in the world, we need to balance openness with the ability to continue building.
Commercial licensing is the way we fund continued development. It's what allows us to keep investing in the next model, maintain the infrastructure, and grow the team. For our enterprise customers, this means you get a stable, production-ready TTS model backed by a dedicated team, rather than relying on unsupported community updates.
We made a deliberate choice: release everything the community and developers need to use, study, and build on S2 for free, and offer commercial licenses for companies that want to deploy it in production. We think that's the right balance for where we are today.
What This Means for Enterprise Customers
If you're evaluating S2 for commercial use, here's what the path looks like:
Evaluate freely. Download the weights, run them on your infrastructure, benchmark against your use cases. The research license covers all of this at no cost.
Commercial licensing is straightforward. When you're ready to ship, reach out to us at business@fish.audio. We offer commercial licenses designed to give companies the flexibility and legal clarity they need to build with confidence. Whether you need API access, on-premise deployment, white-label integration, or a custom arrangement, we'll work with you to find the right structure.
You have full technical control. Because we released the fine-tuning code and inference engine alongside the weights, you can build deep integrations knowing the underlying stack is transparent and inspectable. A commercial license grants you the right to deploy in production.
Why We Keep Releasing What We Can
We believe in being as open as we can sustain. That's why we released the full inference engine when we could have kept it proprietary. That's why we published the complete technical report. That's why the fine-tuning code ships alongside the weights.
At the same time, the community has always been at the core of Fish Audio. Fish Audio started as an open-source project. 6M creators and 2M+ voice models on our platform didn't happen because of us. It happened because of this community. That's why we keep opening what we can, and why we're not going anywhere.
Try S2: fish.audio/s2
GitHub: github.com/fishaudio/fish-speech
HuggingFace: huggingface.co/fishaudio/s2-pro
Commercial licensing: business@fish.audio
