/Tech10d ago

Boson AI releases Higgs Audio v3, a 4B-parameter chat-native TTS model that streams audio before sentences are fully generated

The model supports zero-shot voice cloning across 100 languages

--0--

#290

Original post

Banghua Zhu#1718

LMSYS Org@lmsysorg

🎉 Meet Higgs Audio v3 TTS from @boson_ai, a ~4B chat-native TTS model for real-time voice agents. Day-0 support is live in SGLang-Omni!

> Low-latency streaming > 100 languages, single-digit WER/CER > Zero-shot voice cloning from a short clip > 20+ inline tokens for emotion, style & SFX > 14.74 req/s @ RTF 0.262 on 1× H100

👉 Cookbook: http://sgl-project.github.io/sglang-omni/cookbook/higgs_tts.html Run it now with SGLang-Omni!

BosonAI@boson_ai

Higgs Audio v3 TTS is here.

Built for voice AI that speaks, not just reads: • 100 languages with single-digit WER/CER • inline control over emotion, style, prosody, and sound effects • API, Workspace, and open weights • Blog 👉 https://www.boson.ai/blog/higgs-audio-v3-tts

Watch the demo 👇

2:14 PM · Jun 4, 2026 · 7.4K Views

Sentiment

Users are excited about Boson AI and LMSYS releasing the open-weights Higgs Audio v3 TTS model because it seems faster than the prior 8B version.

Pos

100.0%

Neg

0.0%

1 comments with sentiment.

Cluster Engagement

Views

Comments

Reposts

Bookmarks

Expand data

Posts from X

Most Activity

VIEWS4.3KBOOKMARKS46LIKES70RETWEETS6REPLIES6

LMSYS Org@lmsysorg

🎙️ New blog: Higgs Audio v3 TTS on SGLang-Omni: Real-Time, Controllable Speech for Voice Agents

Higgs (@bosonai) is a ~4B chat-native TTS model (Qwen3-4B backbone, interleaved text + audio tokens) built for streaming voice: 🎙️ Synthesis starts before a full sentence arrives 🌍 100 languages at single-digit WER/CER 🗣️ Zero-shot voice cloning from a short reference clip, across languages 🎛️ 20+ inline control tokens for emotion, style, prosody & sound effects

⚡ Serving with SGLang-Omni Runs as a multi-stage pipeline: preprocessing → audio encoder → TTS engine → vocoder ✅ CUDA Graph decode ✅ Async one-step-lookahead ✅ Batched vocoder & encoder ✅ RadixAttention partitioned by reference audio to amortize repeated cloning

📊 Performance: 14.74 req/s at RTF 0.262 with 16 concurrent requests on a single H100 (bf16)

10d4.3K7046

Alex Smola@smolix

Good voice agents need speech that sounds human, keeps up with a real conversation, and works in your language, on affordable hardware. Today @boson_ai and the @lmsysorg SGLang team release Higgs Audio v3, an open 4B text-to-speech model that hits all three.

10d89992

Alex Smola@smolix

What Higgs Audio v3 does: - voice cloning from a short clip, reusable across languages - 100+ languages at single-digit error rates - emotion, style, prosody set inline via control tokens - streaming, so it speaks before the sentence is finished

Alex Smola@smolix

10d41931

LMSYS Org@lmsysorg

🎉 Meet Higgs Audio v3 TTS from @boson_ai, a ~4B chat-native TTS model for real-time voice agents. Day-0 support is live in SGLang-Omni!

> Low-latency streaming > 100 languages, single-digit WER/CER > Zero-shot voice cloning from a short clip > 20+ inline tokens for emotion, style & SFX > 14.74 req/s @ RTF 0.262 on 1× H100

👉 Cookbook: http://sgl-project.github.io/sglang-omni/cookbook/higgs_tts.html Run it now with SGLang-Omni!

BosonAI@boson_ai

Higgs Audio v3 TTS is here.

Watch the demo 👇

10d7.4K11669

Alex Smola@smolix

The other half of the release is serving. A TTS model like this is a multi-stage pipeline, and making it fast end to end is the real work. With SGLang-Omni it runs faster than real time on a single H100: ~15 req/s, RTF 0.26, ~62 audio-seconds generated per second.

Alex Smola@smolix

10d48010

Alex Smola@smolix

Open weights, open serving. If you are building something that talks back, this is ready to use.

Model: https://huggingface.co/bosonai/higgs-audio-v3-tts-4b Serving (SGLang-Omni): https://github.com/sgl-project/sglang-omni

10d31

Barathwaj Anandan@BarathAnandan7

@lmsysorg @boson_ai Oooo i love this. Seems faster than the other 8B release yesterday. Couldn’t get it to be realtime 👀👀

10d12