Higgs TTS 3 does 102 languages — but read the licence before you sell anything

Short answer: Higgs TTS 3 is 4.65 billion parameters in a single 9.31 GB safetensors file, covers 102 languages (85 of them at production quality), and is licensed for research and non-commercial use — with a Creator Use Grant that lets you monetise podcasts and videos for free as long as you credit Boson AI's Higgs Audio. You cannot host it as an API, resell it, or embed it in a product without a separate commercial licence.

That middle clause is the whole reason this page exists. Boson AI added the Creator Use Grant on 9 July 2026 in the same commit that renamed the model from "Higgs Audio v3 TTS" to "Higgs TTS 3", and almost nobody has written it up. If you narrate YouTube videos, run a podcast, or produce audiobooks, the licence question you have been avoiding has a written answer — and it is more generous than you would guess from the words "Research and Non-Commercial" at the top.

Everything below is read off the official model card and repository listing, not from memory. Where a number is Boson's own claim rather than something measured on consumer hardware, it says so.

---

The licence, in one table

The licence is not a single yes or no — it splits by what you are doing with the model, not by whether you make money. That distinction is what trips people up.

The credit requirement has teeth in one specific way: the model card says the acknowledgment must appear either in the audio itself or "in the accompanying text, displayed prominently" — and adds that it must be clearly visible and not hidden at the bottom of the credits or annotations. A line at the top of your video description satisfies it. A tag buried under forty hashtags does not. Boson suggests this exact string:

This audio was created with Boson AI's Higgs Audio — https://www.boson.ai/higgs-audio

Beyond the commercial split, the use restrictions are absolute and survive the Creator Use Grant: no non-consensual voice cloning, no impersonation, no fraud, no election deception, no biometric surveillance, and AI-generated audio must be disclosed where the law requires it.

If licence terms are the thing you are actually shopping for, our XTTS and Coqui commercial licence breakdown covers the same question for the model most people reach for first — and the answer there is less friendly.

---

What VRAM you actually need

The only number we can verify is 9.31 GB — that is the size of model.safetensors in the repository, holding 4,654,850,537 parameters in BF16. Boson publishes no VRAM table and no official quantised build, so anything beyond that file size is headroom you should measure on your own box rather than take from a blog post. Here is the honest per-tier read:

The architecture explains why this stays modest. Per the model card the backbone is a ~4B autoregressive decoder (36 layers, hidden size 2560, GQA 32/8), and the training sequence length — and therefore the context window — is 8,192 tokens. There is no 128K KV cache to budget for. Audio is encoded into 8 codebooks at 25 frames per second and decoded back to 24 kHz waveform.

For a wider comparison of what voice models cost in memory, see our roundup of the best local TTS models.

---

The 102 languages, tiered

Boson claims single-digit WER/CER across 102 languages, and — unusually — publishes the split rather than rounding it into one number. 85 languages land under 5 WER/CER and are labelled polished and production-quality. 17 land between 5 and 10 and are labelled usable but less polished.

The 17 in the softer tier: Albanian, Chichewa/Nyanja, Eastern Punjabi, Ganda, Icelandic, Irish, Kabyle, Kabuverdianu, Kamba, Latin, Luxembourgish, Oromo, Pashto, Sindhi, Somali, Umbundu and Welsh.

Everything else — English, Chinese, Spanish, French, German, Italian, Japanese, Korean, Portuguese, Russian, Hindi, Arabic, Bengali, Tamil, Telugu, Vietnamese, Indonesian, Turkish, Polish, Dutch, Swedish, Swahili, Zulu and the rest of the list — sits in the under-5 group.

Two honest caveats. First, WER and CER measure intelligibility, not whether a native speaker finds the accent convincing; those are different tests and only the first one is published. Second, these are Boson's own evaluations. If you are building a channel in Welsh or Icelandic, generate a two-minute sample and listen before you commit.

---

Setup

There are two supported serving paths, both exposing an OpenAI-compatible /v1/audio/speech endpoint — which means whatever you already wrote against OpenAI TTS mostly just works after a base-URL swap.

SGLang-Omni (Boson's primary recommendation)

docker pull lmsysorg/sglang-omni:dev
docker run -it --gpus all --shm-size 32g --ipc host --network host --privileged \
  lmsysorg/sglang-omni:dev /bin/zsh

git clone https://github.com/sgl-project/sglang-omni.git && cd sglang-omni
uv venv .venv -p 3.12 && source .venv/bin/activate
uv pip install -v -e .

Then pull the weights and serve:

export HF_TOKEN=hf_xxxxxxxxxxxxxxxx
hf download bosonai/higgs-tts-3-4b

sgl-omni serve --model-path bosonai/higgs-tts-3-4b --port 8000

vLLM-Omni (the simpler alternative)

hf download bosonai/higgs-tts-3-4b

vllm-omni serve bosonai/higgs-tts-3-4b \
  --host 0.0.0.0 --port 8095 \
  --trust-remote-code --omni

First synthesis

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello, how are you?"}' \
  --output output.wav

Voice cloning

Zero-shot cloning takes a reference clip. The model card is explicit that supplying the reference transcript alongside the audio "materially improves cloning fidelity" — this is the single most common thing people leave out.

import requests

resp = requests.post(
    "http://localhost:8000/v1/audio/speech",
    json={
        "input": "Have a nice day and enjoy the sunshine.",
        "references": [{
            "audio_path": "ref.wav",
            "text": "Hey, Adam here. Let's create something that feels real.",
        }],
        "temperature": 0.8, "top_k": 50, "max_new_tokens": 1024,
    },
)
open("output.wav", "wb").write(resp.content)

Set "stream": true to get base64-encoded WAV chunks back over server-sent events as the vocoder emits them; Boson describes this as sub-second time-to-first-audio. For a real-time assistant that is the difference between usable and not. If you are wiring this into a longer pipeline, our local voice cloning walkthrough covers the reference-audio prep that determines most of your output quality.

---

Speed: the H100 numbers and what they mean

Every throughput figure Boson publishes was measured on a single H100 — you should read them as a shape, not as a promise about your 4070.

The benchmark is Seed-TTS EN, full set, N=1088 per run, BF16, CUDA graphs on, max_running_requests=16, each row the mean of three runs:

The useful reading is the RTF column. At concurrency 1 the H100 generates audio roughly 6.8x faster than real time (RTF 0.147), and even saturated at 16 concurrent requests it is still about 3.8x faster than real time per request. A consumer card will be slower — how much slower depends on your memory bandwidth and whether your runtime enables CUDA graphs — but the headroom is large enough that a 4090 or even a 4070 should still comfortably beat real time on a single stream. We have not measured that ourselves, and we are not going to invent a number for it.

What that means practically: a 20-minute narration script is minutes of generation, not hours, on anything modern. Compare that with the timings in our audiobook generation guide if you are planning long-form work.

---

Inline control tokens

This is the feature that separates Higgs TTS 3 from the read-the-text-flatly generation of local TTS. All tags use <|category:value|> syntax and can be dropped mid-sentence.

• Emotion — 21 tokens, from <|emotion:elation|> and <|emotion:amusement|> through <|emotion:contemplation|>, <|emotion:bitterness|> and <|emotion:helplessness|>.
• Style — <|style:singing|>, <|style:shouting|>, <|style:whispering|>.
• Sound effects — <|sfx:laughter|>, <|sfx:sigh|>, <|sfx:cough|>, <|sfx:crying|>, <|sfx:sneeze|> and more. The card advises pairing each with the matching onomatopoeia immediately after it, so <|sfx:laughter|>Haha rather than the tag alone.
• Prosody — speed from about 0.65x to 1.4x, pitch at roughly -3 to +2.5 semitones, plus <|prosody:pause|> (≈400-700 ms) and <|prosody:long_pause|> (≈700-1500 ms), and expressiveness up or down.

Those pause tokens are quietly the most useful thing here. Getting a narrator to breathe in the right places is normally an editing job; here it is a tag.

---

Where it is better and worse than XTTS-v2

Better: language coverage is not close — 102 languages against XTTS-v2's short European-and-Asian shortlist. The licence position for creators is materially clearer, because the Creator Use Grant is written down rather than inferred. Inline emotion, style, sfx and prosody control is a category XTTS simply does not have. And an 8,192-token context with streaming makes it a plausible voice-agent backend rather than only a batch narrator.

Worse: it is bigger. 9.31 GB of BF16 weights against XTTS-v2's much smaller footprint means the 8 GB tier goes from routine to experimental. Tooling is younger — SGLang-Omni and vLLM-Omni are the supported paths, and neither is as widely deployed as the Coqui/XTTS ecosystem with its years of forks, GUIs and tutorials. There is no official quantised build. And for a product you intend to sell, both models land you in the same place: a conversation with the vendor.

If you want the older, smaller, better-documented option, our XTTS-v2 voice cloning guide and the Kokoro vs XTTS vs Chatterbox comparison cover that side, and F5-TTS is worth a look if you want a lighter model and fewer moving parts. The Coqui TTS model page has the specs side by side.

---

Honest limitations

• 8,192-token context. That is the training sequence length, so very long single requests need chunking. Fine for narration, a constraint for anything that wants one enormous prompt.
• No official quantisation. Every sub-BF16 option — 4-bit AWQ, MLX 4-bit and 6-bit, NVFP4, ONNX, GGUF — is community work. At our mid-August check the busiest community port was an MLX BF16 conversion at roughly 1,200 downloads and the quantised ones sat in the low hundreds, against ~218,000 for the official weights. Orders of magnitude less scrutiny, and the ranking will have shifted by the time you read this.
• The credit requirement is a real obligation. If you forget it on one video, you are outside the Creator Use Grant for that video. Build it into your description template rather than remembering it each time.
• Benchmarks are self-reported. The WER/CER tiers, the win-rate tables and the throughput numbers all come from Boson AI, run by Boson AI. That is normal, and it is still worth saying out loud.
• Consent is not optional. Non-consensual cloning and impersonation are prohibited by the licence regardless of what you are doing with the output. A model this good at zero-shot cloning makes that easy to violate carelessly.

---

Verdict

If you are a creator, this is the most interesting local TTS release of the summer, and the licence is the reason — not the model card's headline numbers. When we checked the repository in mid-August 2026 it was running around 218,000 downloads in a rolling 30-day window with roughly 720 likes — figures that move daily, so treat them as a snapshot rather than a scoreboard. Independent adoption in audio.cpp is the more durable signal that it works outside Boson's own stack.

Get it if: you narrate content in more than one language, you want emotion and pause control without post-production, you have 12 GB of VRAM or more or an Apple Silicon Mac, and you are willing to put one line of credit in your video description.

Skip it if: you are building something you sell — a hosted API, an app, a service. In that case the Creator Use Grant does not reach you, and you should either contact Boson AI for commercial terms or start from a permissively licensed model instead.

---

FAQ