Audio8-TTS-Realistic-0.6B vs Qwen3-TTS

Official Audio8-TTS-Realistic-0.6B banner from the Edge0 Hugging Face card — Edge0/Audio8-TTS-Realistic-0.6b card audio8-banner.jpeg. Source · License: Apache-2.0.

Official Qwen3-TTS introduction figure from the Qwen3-TTS repository assets — Qwen3-TTS-Repo/qwen3_tts_introduction.png (CustomVoice README). Source · License: Qwen product image.

Verdict. Pick Audio8-TTS-Realistic-0.6B if you want Edge0’s ungated 0.6B DualAR checkpoint (23 languages, bundled 44.1 kHz codec, trust_remote_code). Pick Qwen3-TTS if you want Qwen’s Apache-2.0 12Hz family (CustomVoice, VoiceDesign, Base 0.6B/1.7B). They are not the same weights. Language lists differ. Vendor Seed-TTS tables stay vendor. We have no sourced bake-off.

Compare Audio8-TTS-Realistic-0.6B and Qwen3-TTS.

Dimension	Audio8-TTS-Realistic-0.6B	Qwen3-TTS	Source
Job	Single DualAR TTS checkpoint (zero-shot clone or no-reference)	Family (CustomVoice / VoiceDesign / Base + 12Hz tokenizer)	Official cards
License	Apache License 2.0 (LICENSE on the weight repo)	apache-2.0 card tag; Apache-2.0 LICENSE on QwenLM/Qwen3-TTS (CustomVoice blob 404)	Hugging Face / GitHub LICENSE files
Gating	No	No (cards we opened)	Hugging Face API
Developer	Edge0	Qwen	Official cards
Parameters	601,159,424 excluding codec	0.6B and 1.7B cards (1.7B CustomVoice safetensors 1,916,676,352)	Official READMEs / HF API
Languages (card)	23 (Preview 11 + 12 added)	10 major languages on the family README	Official READMEs
Codec / sample rate	Bundled 44.1 kHz, 10 codebooks, ~21.5 frames/s	Qwen3-TTS-Tokenizer-12Hz (family README)	Official READMEs
Hugging Face likes (lead)	207 (Realistic, 2026-11-02)	2,191 (1.7B CustomVoice, 2026-11-02)	Hugging Face API
GitHub stars	2,238 (Edge0-AI/Audio8_TTS)	14,154 (QwenLM/Qwen3-TTS)	GitHub API (2026-11-02)
Quality vs ElevenLabs	Not independently tested	Not independently tested	n/a

When to choose which

One 0.6B DualAR file + bundled 44.1 kHz codec: Audio8-TTS-Realistic-0.6B — Preview parent is a different card. Quality still varies by language.
Named speakers / VoiceDesign instruct: Qwen3-TTS — CustomVoice and VoiceDesign are separate 12Hz cards. Not an ElevenLabs studio.
3-second voice clone as a Base checkpoint: Qwen3-TTS — 0.6B-Base and 1.7B-Base READMEs. Audio8 also does zero-shot clone; different API.
Same Apache-2.0 grant: Tie — Audio8 ships LICENSE on the weight repo. Qwen cards tag apache-2.0; GitHub LICENSE is Apache-2.0. Not legal advice.

Read the entity pages: Audio8-TTS-Realistic-0.6B and Qwen3-TTS. Both hang off ElevenLabs alternatives. Related Liquid audio (any-to-any, not a DualAR TTS file): LFM3-Audio-2B. Realistic-0.6B ≠ Qwen3-TTS.

Comparison graphic: Audio8-TTS-Realistic-0.6B DualAR 0.6B / 23 languages versus Qwen3-TTS 12Hz CustomVoice / VoiceDesign / Base family under Apache-2.0 — forkalt comparison graphic (2026-11-02). Not a ranking.

Why this comparison exists

People searching compact open TTS hit Audio8 Realistic 0.6B and Qwen3-TTS in the same cluster. Audio8’s own card even cites Qwen3-TTS Seed-TTS numbers in a reference table and says that is not a matched ranking. This page keeps the artifacts distinct with sourced rows only.

Drama 3 stays on Drama 3 (35B MoE, Community License). Phonon stays on Phonon (31M Apache-2.0 English). Those are different sizes and licenses.

Sourced snapshot (2026-11-02)

Field	Audio8-TTS-Realistic-0.6B	Qwen3-TTS (family)
Host	huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b (not gated)	Lead: huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice (not gated)
HF likes / downloads	207 / 4,613	2,191 / 1,749,479 (CustomVoice)
License	Apache-2.0 LICENSE on the card	apache-2.0 tags; GitHub Apache-2.0
Size	0.6B DualAR (codec extra)	0.6B and 1.7B 12Hz cards
Job	One TTS checkpoint	CustomVoice / VoiceDesign / Base
Languages	23 listed	10 listed on the family README
Official local path	Transformers + trust_remote_code; SGLang Omni named	`pip install -U qwen-tts`
Quality vs ElevenLabs	Not independently tested	Not independently tested

License

Audio8 ships Apache License 2.0 on the weight repo. Qwen cards tag apache-2.0; the CustomVoice blob was 404 on 2026-11-02; QwenLM/Qwen3-TTS LICENSE is Apache License 2.0. Audio8’s card still asks for consent before cloning a real voice. Not legal advice.

How to choose

Need one 0.6B DualAR file / 23 languages: Audio8-TTS-Realistic-0.6B.
Need Qwen CustomVoice, VoiceDesign, or Base: Qwen3-TTS.
Need 35B directed-performance TTS (Community License): Drama 3.
Need on-device 31M English Apache-2.0: Phonon.
Need Liquid audio any-to-any (ASR+TTS+tools, LFM 1.0): LFM3-Audio-2B. Omni (adds vision): LFM3-Omni-2B.
Need a pixel UI, not TTS: Fooocus, Iris-3B vs Qwen-Image-2.1-Turbo.
Need chat weights: Nemotron 3.5 Bolt, LFM3-1.7B.
Need style rewrite: humanizer.

See also Video & Audio and Models. Catalog: home. Editorial: About. Corrections: contact.

FAQ

Which one should I download first?

Need one DualAR 0.6B file with 23 listed languages — Audio8-TTS-Realistic-0.6b (trust_remote_code). Need Qwen CustomVoice, VoiceDesign, or Base — start from the Qwen3-TTS overview and pick the variant. They do not substitute for each other.

Did Audio8 beat Qwen3-TTS on Seed-TTS?

We do not claim that. Audio8’s card prints a reference table that cites Qwen3-TTS report numbers and says it is not a strictly matched ranking. Qwen’s 1.7B-Base Seed-TTS row (test-zh 0.77 / test-en 1.24) is a different vendor table. forkalt has not re-run either bench.

Why this pairing instead of Drama 3?

Searchers looking for compact Apache-2.0 TTS hit both the Realistic 0.6B name and Qwen3-TTS. Drama 3 is a 35B MoE under a Community License with a USD 500,000 commercial threshold. It stays on the ElevenLabs hub and its own model page.

Does either replace ElevenLabs?

No sourced bake-off. Both hang off ElevenLabs alternatives as downloadable TTS weights, not hosted studios.

Sources & verification

Last verified 2026-11-02 · Last updated 2026-11-02· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.