Qwen3-TTS
Qwen · tts · reviewed by Stan K. · Source: data snapshot · data as of 2026-11-12
Apache 2.0 weights with a confirmed 3-second rapid voice clone — the commercial-friendly cloning pick if you don't want a paid API.
#4 of 16 in Tts
ToolScore breakdown — why 82/100
10 criteria × 10 points. Low pillars are visible on purpose — that's the point of a score you can audit. Methodology →
What it does well
- ▸3-second rapid voice clone — confirmed, not marketing: the headline feature works
- ▸The commercial-friendly cloning pick if you don't want a paid API — Apache 2.0, no NC trap
- ▸Cloning runs on your hardware: no vendor cloning policy, no per-clone billing, no policy flip can take it away
- ▸Params published (0.6B / 1.7B) — you can size the deployment class before you commit
Where it costs you
- ▸VRAM requirements not published in our snapshot — verify pending; the watts bill is uncomputable until then
- ▸No latency number published — nothing to check generation speed against before you deploy
- ▸No row in the AA Speech Arena table for the local model — quality rests on the confirmed clone, not an Elo rank
Hidden traps — the stuff the pricing page doesn't say
Our hub publishes VRAM requirements only when a source confirms them; for Qwen3-TTS the cell stays "—". The 0.6B / 1.7B parameter counts say small-to-mid GPU, but budgeting exact watts — the local cost model — is verify pending.
Know the limits before you pay: read the traps above, then decide with eyes open.
Affiliate link where available — never affects the score.
- internal ToolScore TTS hub — local table: Apache 2.0 license, 0.6B/1.7B params, 3-second clone confirmed (licenses from LICENSE files and model cards) verified
Data as of 2026-11-12. Prices and limits change monthly — report anything stale. How we verify: methodology.
Closest alternatives
Top-5 quality (Elo 1296) with the most honest meter in the business — the official ~750–800 credits/
≈$0.014/min with the only published token rate in the table — 25 tokens/sec makes the math checkable
MIT-licensed and CPU by design — the lightweight workhorse for the machine where there is no GPU at
Qwen3-TTS review
In-depth analysis · last updated 2026-11-12
The clone you can keep
Qwen3-TTS is the second-ranked local model in our hub, and its case fits in one line: a 3-second rapid voice clone — confirmed — under an Apache 2.0 license. That combination does not exist anywhere else in the local table. Paid APIs sell cloning (Cartesia locks it away from its free tier entirely); open models that allow cloning usually carry an NC trap. Qwen3-TTS gives you the clone on your own hardware, commercially usable, with no vendor policy standing between you and the voice.
The economics — free, with one blank cell
The cost model is the standard local one: watts, not minutes. Params are published (0.6B / 1.7B), which puts the deployment class in reach — but the VRAM cell in our hub stays "—" (verify pending), so the one number the local cost model needs, the watts bill, is the one we cannot compute. Against ≈$0.14–0.15/min subscription credits, even a mid-GPU deployment wins on volume; against Kokoro's CPU-only anchor, Qwen3-TTS will cost real watts.
Real limits
Structurally there are none: self-hosted means no character meter, no clone-per-month cap, no concurrency ceiling. The blank cells are the limit — no VRAM figure, no latency number, no AA Speech Arena row for the local model. What is confirmed is the thing that matters most: the clone works, in 3 seconds.
How the pillars land
Price/Value 9: electricity only, docked one point because the hardware bill is unpublished. Quotas 9: self-hosted, no meter exists. Latency sits at the neutral 5 — no latency fact exists in our snapshot, and we don't publish guesses. Training use & voice cloning 9: cloning confirmed on your own hardware, and local inference keeps text on your box; no vendor training-use policy applies. API & self-host 10: Apache 2.0, the commercial-friendly license. Feature set 9.5: the confirmed 3-second clone is the strongest single feature in the local table. Track record 8: the clone is verified working; the project itself has no long-run history in our snapshot. Ease 9.5: cloning without a paid API — no signup, no keys, no policy console. Pricing transparency 9: the license is published; the missing VRAM figure costs the point. Support & docs is the neutral 5 — no support facts exist for it in our snapshot.
Verdict — COMMERCIAL CLONING PICK, 82/100
If voice cloning is the requirement and a paid API is what you're trying to avoid, this is the pick: Apache 2.0, confirmed 3-second clone, no vendor in the loop. Size the hardware before you commit (the VRAM cell is still blank in our table), and pair it with Kokoro for the non-cloning bulk — the two Apache 2.0 models cover each other's gaps.