Together AI added hosted Qwen3 TTS on November 20, 2026, with nine preset voices, plain-language delivery instructions, ten languages per voice, and pricing at $20 per 1 million characters. The company says its inference team tuned serving so first audio can start in as little as 97ms, aiming at live assistants rather than batch narration alone.
The launch sits next to Together Storage Accelerator and agent eval work as Together pitches the full stack, not only raw GPUs. For builders comparing open TTS, see Nari Labs' Qwen3-TTS serving write-up, Hugging Face speech-to-speech defaults, and Zyphra's ZONOS3-Mini release.
TL;DR: what shipped and what it costs
| Question | Answer |
|---|---|
| Model id | Qwen/Qwen3-TTS-12Hz-1.7B per Together TTS docs |
| Price (Together list) | $20 / 1M characters |
| Voices | Nine presets (vivian, serena, ryan, aiden, …) |
| Languages | 10 (zh, en, ja, ko, de, fr, ru, pt, es, it) |
| Style control | extra_params.instructions (plain language, up to 500 chars) |
| Latency claim | As low as 97ms to first audio (Together launch post) |
| Timestamps | Word alignment not supported on this model |
What Together announced
In its November 20 launch post, Together frames Qwen3 TTS as separating who speaks from how they speak. You pick one of nine preset voices. The voice stays fixed while you change delivery with a short instruction such as excited and upbeat, or move the same voice across languages.
Together says the model is available on serverless inference and dedicated endpoints. The post highlights three demo angles: one English line read with three different instructions, one voice reading an order update in four languages, and latency tuned for real-time conversation.
Authors credited on the post include Anthony Ji, Charles Duong, Sebastien Beurnier, Sahil Yadav, and Sonny Khan. Pricing and API fields below come from Together's documentation, which Together updated to list Qwen3 alongside Orpheus and Kokoro.
Speech waveform lines illustrating Qwen3 TTS multilingual output on Together AI
Voices, languages, and instructions
Together documents nine preset voices. Each has a native language but can speak every supported locale when you set language:
| Voice | Native language (per Together docs) |
|---|---|
| vivian, serena, uncle_fu, dylan, eric | Chinese variants |
| ryan, aiden | English |
| ono_anna | Japanese |
| sohee | Korean |
Supported language codes include zh, en, ja, ko, de, fr, ru, pt, es, it. That matters for products that want one branded voice across European and Asian locales without training a new speaker embedding per market.
Style instructions ride in extra_params.instructions. Together's docs show a congratulatory line with Speak warmly and happily. The same mechanism works with stream: true and on the WebSocket API when you set extra_params.instructions in tts_session.updated or on the connection URL.
Qwen3 TTS also supports an optional seed in extra_params for best-effort reproducibility when input, voice, and seed match.
Pricing in context on Together
Together's public price table in its TTS docs (fetched November 20, 2026) lists:
| Model on Together | Price per 1M characters |
|---|---|
| Kokoro | $4 |
| Orpheus 3B | $15 |
| Qwen3 TTS 1.7B | $20 |
That positions Qwen3 as a mid-tier hosted option: cheaper than Cartesia Sonic 2 at $65 on the same table, more than Kokoro, with instruction steering and multilingual presets as the differentiator.
Teams already optimizing Qwen3-TTS on owned hardware should still read Nari Labs' speed-cost frontier post, which reported sub-50ms p95 time-to-first-audio on one H100 with a custom stack. Together's 97ms figure is a hosted claim on Together's stack, not a universal ceiling for the model weights.
How to call the API
Together documents three delivery modes: standard HTTP, streaming HTTP for low time-to-first-byte, and WebSockets for interactive agents. Qwen3 supports wav, mp3, raw, and mulaw response formats. Sample rates are 24000 Hz default or 8000 Hz.
Example curl from Together docs (API key via environment):
curl -X POST "https://api.together.ai/v1/audio/speech" \
-H "Authorization: Bearer $TOGETHER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen3-TTS-12Hz-1.7B",
"input": "Congratulations on the new job! I am really proud of you.",
"voice": "ryan",
"language": "en",
"response_format": "wav",
"extra_params": {"instructions": "Speak warmly and happily."}
}' \
--output speech.wav
For voice agents, Together points to a real-time STT plus TTS pipeline guide in the same docs set. Pair this endpoint with GPT Realtime voice APIs or local STT from superwhisper S1-mini depending on whether the LLM leg stays on Together or elsewhere.
API adapter diagram for hosted text-to-speech requests on Together AI
When to pick Together Qwen3 versus local or other APIs
Pick Together Qwen3 TTS when you want Qwen's instruction-steered presets without operating GPUs, you already bill inference to Together, or you need HTTP and WebSocket speech from one vendor next to Llama, Qwen LLMs, and GPU clusters.
Self-host or use another host when cost per million characters dominates at your volume, you need word timestamps, or you must keep audio on-prem for compliance. Voicebox-style local studios and Miso One remain relevant for teams optimizing capex over opex.
Compare hosted rivals using HeyGen on OpenRouter pricing claims and DeepL Voice Aure tone control when emotional range and translation matter more than open weights.
Streaming HTTP versus WebSocket for voice agents
Together's TTS docs draw a bright line between transport choices:
| Mode | Best for | Qwen3 notes |
|---|---|---|
| Standard HTTP | Batch files, voicemail renders, offline clips | Returns complete wav or mp3 |
| Streaming HTTP | Low time-to-first-byte on full sentences | Supports stream: true with instructions |
| WebSocket API | Chatbots that receive tokens incrementally | Send partial text; audio chunks flow out |
For a voice agent shaped like Hugging Face's speech-to-speech recipe, you might keep STT and the LLM on one stack and add Qwen3 on Together only for the TTS leg. That avoids shipping another GPU image just to test Alibaba's open weights.
If you already run Parakeet TDT plus a hosted LLM, swapping the TTS endpoint to Together is mostly a model string change plus voice name. Test language codes on every locale you ship; Together documents lowercase locale tags such as zh-hk for Cantonese variants on other models, while Qwen3 lists ten core languages.
Cost math for product owners
At $20 per million characters, rough napkin math:
| Monthly spoken output | Character ballpark | Together Qwen3 bill (list price) |
|---|---|---|
| Light assistant | 500k chars | ~$10 |
| Consumer app with daily briefings | 5M chars | ~$100 |
| Call-center narration prototype | 50M chars | ~$1,000 |
Compare against GPU rental if you self-host Qwen3-TTS 1.7B: you pay electricity, ops, and engineering instead of the per-character meter. Nari Labs claimed about $2 per million characters at 10 RPS on one H100, but that assumed their custom stack, not your team's iteration time.
Kokoro on Together at $4 remains the budget tier when you only need intelligibility, not instruction steering across ten languages.
What people are asking after the launch
Do I need a dedicated endpoint? Together says serverless works day one; dedicated endpoints matter when you have steady QPS and want isolation. The launch post mentions both.
Can I clone a custom voice? Qwen3 on Together is preset voices only in the docs we read. Custom cloning paths live in other models (Orpheus voice names, Kokoro mixing) on the same platform.
Is this the same as Qwen3-TTS on Hugging Face? The weights family matches Qwen3-TTS-12Hz-1.7B, but Together adds its serving tuning and billing wrapper. Benchmark their TTFA in your region rather than trusting a single blog number.
Will instructions work in every language? Together demos show aiden reading English with emotional instructions, then the same voice in other languages. You still pass language explicitly; instructions are not a substitute for locale tags.
Limits and honest caveats
Together's docs are explicit about Qwen3 gaps:
- No word-level alignment on this model.
- No opus, aac, or flac output formats (unlike some Minimax models on the same platform).
- Seed reproducibility is best-effort only.
The 97ms latency line is marketing copy on Together's blog until you benchmark your region, sentence length, and concurrent streams. For production voice agents, budget LLM time-to-first-token separately; as Hacker News commenters noted on Nari's post, TTS TTFA alone does not cap end-to-end conversational delay.
Compliance and data handling checklist
Together hosts customer text on inference. For regulated apps:
- Log retention policies in your DPIA even when Together does not train on API traffic by default contract wording you sign.
- Separate PII from prompts where possible; instructions fields are as sensitive as the spoken line.
- Document voice choice in accessibility reviews; emotional instructions can skew perceived tone in banking or healthcare bots.
None of that is unique to Qwen3, but instruction-steered TTS makes marketing copy generation tempting — legal teams may want review gates before auto-speaking dynamic CRM fields.
Background: why Qwen3-TTS mattered before Together hosted it
Alibaba's Qwen team open-sourced Qwen3-TTS weights earlier in 2026, and explainx.ai already tracked serving optimizations (Nari Labs), local CPU stacks (Kokoro guide), and multi-model studios (Voicebox). Together's launch is the managed API layer: same family, vendor SLO, and a single invoice next to Llama or DeepSeek calls.
That matters for teams who standardized on Together during agent eval pushes or prefix-cache experiments and do not want a second GPU vendor only for speech.
Deep dive on instructions: treat them like a secondary prompt. Short, testable phrases beat paragraphs. Start with three templates (neutral, urgent, empathetic) per product surface and A/B listener studies before letting users type free-form instructions in production.
Related reading
- Together Storage Accelerator
- Together agent evals at 97% completion
- Nari Labs Qwen3-TTS GPU serving
- Hugging Face speech-to-speech pipeline
- Kokoro local CPU TTS guide
- OpenAI GPT Realtime voice models
- Zyphra ZONOS3-Mini open TTS
Official sources: Together AI — Qwen3 TTS launch · Together AI — Text-to-speech documentation
Pricing, model ids, and latency claims are accurate as of November 20, 2026, per Together's blog and docs.
