# Boson AI (Higgs Audio)

> Boson AI and its open-weight Higgs Audio voice models: Higgs TTS 3's 102 languages, its creator-only license, hosted API, and rivals like ElevenLabs.

Source: https://metavert.io/boson-ai  
Published: 2026-10-21  
Updated: 2026-10-21

**Boson AI** is a Santa Clara, California startup that builds voice and audio models, best known for **Higgs Audio**, a family of open-weight [text-to-speech](https://metavert.io/text-to-speech) models. Its current flagship, Higgs TTS 3, released on June 4, 2026, has about 4 billion parameters, covers 102 languages and clones a voice from a short reference clip, and as of October 2026 it is one of the most capable voice models whose weights anyone can download. Its license is the catch: it allows research, non-commercial use and monetized creator content, but building a product or API on it requires a separate commercial license.

### Company

Boson AI was founded around 2023 by Alex Smola, a machine-learning researcher who spent seven years at Amazon Web Services as a vice president and distinguished scientist. Fortune reported in July 2026 that the company had raised $70 million, backed by Chinese entrepreneur Su Hua and the venture arm of Singapore's Temasek, and that it was preparing a speech-to-speech model, Higgs RealTime, which Smola said would cost about an order of magnitude less than rivals'. Its stated focus is real-time voice for industries such as finance, telecommunications, healthcare and insurance.

### The Higgs Audio Models

- **Higgs Audio v2 (2025).** A 3-billion-parameter generation model, announced by Boson in July 2025. Its code was released under Apache-2.0 and its weights under a community license based on Meta's Llama 3 license, with a threshold of 100,000 annual active users. Boson recommended 24 GB of GPU memory to run it.
- **Higgs TTS 3 (June 2026).** About 4 billion parameters and 102 languages on the model card, of which Boson reports 85 under 5% word or character error rate when the output is transcribed back. It offers zero-shot voice cloning, 21 emotion tokens, and style and sound-effect tags written into the script.

Boson's naming is inconsistent. The Hugging Face model card and license say "Higgs TTS 3", the GitHub README says "Higgs Audio v3", and the hosted API calls the model `higgs-tts-3`. "Higgs Audio" is the usual name for the family. It is unrelated to [Higgsfield](https://metavert.io/higgsfield), the AI video company.

### License, Hosting and Performance

Higgs TTS 3 ships under the "Boson Higgs TTS 3 Research and Non-Commercial License Agreement". Its Creator Use Grant lets creators publish and monetize podcasts, videos, audiobooks and social posts made with the model, provided they credit Boson AI's Higgs Audio in the audio or prominently in text. The grant does not cover hosting the model for others, offering it through an API or plug-in, or embedding it in an app or service, and the license bans cloning a voice without consent, impersonation, fraud, election deception and biometric surveillance. In practice, a podcaster can use it with a credit line; a startup cannot put it behind its own voice API without a deal.

Running it is a data-center job. Boson publishes a mean latency of 617 milliseconds at one request and 1,079 milliseconds at 16 concurrent requests on a single NVIDIA H100, served through SGLang-Omni or vLLM-Omni. Boson also runs a hosted API at $0.015 per 1,000 input characters, but its documentation says requests should stay around 300 characters and that longer text is "not yet reliably supported"; word timestamps are offered only for English, Chinese and Spanish.

### How It Compares

Against [ElevenLabs](https://metavert.io/elevenlabs), the commercial leader, Higgs TTS 3 offers downloadable weights, privacy and a much lower hosted price per character, while ElevenLabs offers a commercial license on every paid plan, lower latency, long requests and a full platform for dubbing, transcription and music. Against other open models such as Chatterbox, Dia2 and Kokoro, which use permissive MIT or Apache licenses, Higgs TTS 3 is larger and broader in languages but more restricted in use. Commercial real-time rivals include [Cartesia](https://metavert.io/cartesia). As with any voice-cloning model, the consent issues covered in [deepfakes](https://metavert.io/deepfakes) apply. See [ElevenLabs vs Higgs Audio](https://metavert.io/compare/elevenlabs-vs-higgs-audio), [open-weight models](https://metavert.io/open-weight-models), [speech and voice AI](https://metavert.io/speech-voice-ai) and [generative audio](https://metavert.io/generative-audio).

## Related Topics

- [ElevenLabs vs Higgs Audio](https://metavert.io/compare/elevenlabs-vs-higgs-audio) — A commercial leader against an open-weight model
- [Text-to-Speech and AI Voiceover](https://metavert.io/text-to-speech) — How neural TTS works
- [ElevenLabs](https://metavert.io/elevenlabs) — The commercial voice leader
- [Cartesia](https://metavert.io/cartesia) — Real-time voice models
- [Open-Weight Models](https://metavert.io/open-weight-models) — Models you can run yourself
- [Speech and Voice AI](https://metavert.io/speech-voice-ai) — The broader field
- [AI Dubbing and Lip-Sync](https://metavert.io/ai-dubbing) — Translating speech in a cloned voice
- [Deepfakes](https://metavert.io/deepfakes) — Consent and voice cloning

## Further Reading

- [Higgs TTS 3 model card](https://huggingface.co/bosonai/higgs-tts-3-4b) — Hugging Face; parameters, languages and benchmarks
- [Higgs TTS 3 license](https://huggingface.co/bosonai/higgs-tts-3-4b/raw/main/LICENSE) — The research and non-commercial license and Creator Use Grant
- [Higgs TTS 3](https://www.boson.ai/blog/higgs-tts-3) — Boson AI, June 4, 2026
- [Boson API docs](https://docs.boson.ai/models/higgs-tts/overview) — Hosted endpoint, limits and pricing
- [higgs-audio on GitHub](https://github.com/boson-ai/higgs-audio) — Boson AI; code and serving
- [Ex-AWS scientist Alex Smola's Boson AI](https://fortune.com/2026/07/27/ex-aws-scientist-alex-smola-boson-ai-speech-to-speech-voice-model/) — Fortune, July 27, 2026