Insight · Market Coverage
Together AI Brings Real-Time Voice to Its Open Inference Stack
Enterprise-Grade Text-to-Speech Powers Next-Generation Voice Agents
Together AI integrates Qwen3 TTS into its serverless platform, delivering real-time voice with 97ms latency across 9 voices and 10 languages for scalable customer experience automation.
Cite this
Together AI has added Qwen3 TTS to its serverless inference catalog, giving developers access to an open text-to-speech model with nine preset voices, 10-language support, and 97ms time-to-first-audio at $20 per 1M characters [1]. The model's architecture separates speaker identity from delivery style, enabling plain-language emotional control that makes it viable for real-time voice agents [1]. With customer support and experience automation ranking as the top generative AI use case among enterprise decision-makers [2], this addition directly targets the agentic deployments organizations are prioritizing over the next 18 months.
What is Covered in this Article
- Qwen3 TTS launch on Together AI serverless inference [1]
- Real-time voice agent viability via 97ms latency and streaming APIs [1]
- Enterprise demand for voice-enabled customer experience automation [2]
- Open inference as an alternative to proprietary model lock-in [3]
- AI platforms market growth context [4]
The News: Together AI launched Qwen3 TTS on its serverless inference platform on November 20, 2026 [1], making the open text-to-speech model available on both Together Serverless Inference and dedicated endpoints [1]. Priced at $20 per 1M characters [1], the model offers nine preset voices spanning English, Chinese, Japanese, and Korean native speakers [1], each supporting 10 languages including German, French, Russian, Portuguese, Spanish, and Italian [1]. Together AI's serving stack delivers first audio in as little as 97ms [1], with both HTTP streaming and a WebSocket API for real-time voice agent applications [1]. A companion VoiceFinder tool at voicefinder.together.ai lets developers preview voices on custom text before committing to a preset [1]. The underlying model is Qwen/Qwen3-TTS-12Hz-1.7B, developed by the Qwen team at Alibaba Cloud [1].
Analysis: This launch is more than a catalog addition. By combining sub-100ms latency, multilingual coverage, and plain-language delivery control in a single open model, Together AI closes a meaningful gap in its stack and positions itself as a credible full-stack inference provider for voice-enabled agentic applications [1]. The timing aligns directly with where enterprise AI spending is heading.
Architecture That Enables Real-Time Agents, Not Just Batch Synthesis
The design choice that makes Qwen3 TTS production-relevant is its separation of speaker identity from delivery style [1]. Developers choose one of nine preset voices, then control tone through plain-language instructions such as 'excited and upbeat' or 'sad and disappointed,' without altering the chosen voice. This decoupling matters for agentic deployments where a consistent brand voice must adapt emotionally across interactions. Paired with HTTP streaming and a WebSocket API for text that arrives in pieces [1], and first audio in as little as 97ms [1], the model crosses the latency threshold required for real-time conversation. Together AI's inference team explicitly tuned the serving stack for this benchmark, signaling that production voice performance was a deliberate engineering priority, not an incidental outcome. Latency metrics like time-to-first-token are a key production monitoring priority for 41.8% of enterprise decision-makers [2], making the 97ms figure a meaningful differentiator in procurement conversations.
Voice Lands at the Center of Enterprise AI Priorities
The use case alignment here is direct. Customer support and experience automation, covering autonomous chatbots, virtual assistants, and service automation, is the leading generative AI application among enterprise decision-makers at 56.5% [2]. Customer experience, defined as automated service, personalized engagement, and support triage, ranks among the top areas where organizations plan to deploy agentic AI in the next 18 months at 48.6% [2]. Both use cases depend on voice as a delivery layer when interactions move beyond text interfaces. Multilingual capability amplifies this value: every Qwen3 TTS preset voice supports all 10 languages [1], allowing a product to maintain a consistent voice identity when it enters a new market. For enterprises managing global customer operations, that consistency reduces the overhead of maintaining separate voice assets per region.
Open Inference as a Full-Stack Alternative
Together AI's strategic positioning rests on offering developers flexibility that proprietary ecosystems do not. Competing frameworks remain primarily optimized for and tightly integrated with OpenAI's proprietary model ecosystem, enforcing model lock-in rather than providing adapters for third-party foundational models [3]. Together AI's approach, which now spans language, code, and voice, lets development teams compose their stack from best-available open models without committing to a single vendor's roadmap. This resonates with how most enterprises actually operate: 51% of enterprise decision-makers favor a balanced mix of in-house and vendor solutions [2], a preference that maps directly to the flexibility Together AI's serverless inference provides. Adding a production-ready TTS capability reinforces the platform's credibility as a complete alternative rather than a partial substitute. The VoiceFinder tool [1] further reduces adoption friction by letting developers evaluate voices on their own content before writing a single line of production code.
Market Scale Provides the Commercial Backdrop
The AI platforms market is on a base-case trajectory to reach $1.46 trillion in 2026, with a 52.5% CAGR projected through 2030 [4]. At that growth rate, the competitive question is not whether developer spend will materialize, but which platforms will capture it. Together AI's rapid expansion of its model library, now covering language, code, and voice, positions it to serve developers who want a single inference provider across modalities. Qwen3 TTS, built on the Qwen/Qwen3-TTS-12Hz-1.7B model from Alibaba Cloud's Qwen team [1], brings a capable open-weight foundation to that library without requiring Together AI to develop the underlying model itself. That asset-light approach to model acquisition, combined with proprietary serving optimization for latency, reflects a platform strategy focused on inference quality and breadth rather than model ownership.
What to Watch
- Voice agent adoption: which enterprise customer experience teams integrate Qwen3 TTS into production deployments in Q1 2027 and at what scale [2]
- Latency competition: how proprietary TTS providers respond to the 97ms benchmark with their own serving optimizations in the coming quarter [1]
- Multilingual expansion: whether Together AI adds language coverage beyond the current 10 to address markets in Southeast Asia and the Middle East [1]
- Open-model ecosystem momentum: how quickly additional open TTS models appear on the Together catalog as the voice agent market develops [3]
- Enterprise hybrid adoption: whether the 51% of decision-makers favoring mixed in-house and vendor solutions translate that preference into Together AI serverless contracts through Q2 2027 [2]
Sources
1. Qwen3 TTS is now available on Together AI, Together, November 2026
2. 2H 2026 AI Platforms Global Enterprise Decision Maker Survey Report, Futurum Research, October 2026
3. Futurum Signal Report | Agentic Orchestration for Enterprise - November 3, 2026, Futurum Research, November 2026
4. 2H 2026 AI Platforms Market Sizing & Five-Year Forecast, Futurum Research, November 2026
Declaration of generative AI and AI-assisted technologies in the writing process: This content has been generated with the support of artificial intelligence technologies. Due to the fast pace of content creation and the continuous evolution of data and information, The Futurum Group and its analysts strive to ensure the accuracy and factual integrity of the information presented. However, the opinions and interpretations expressed in this content reflect those of the individual author/analyst. The Futurum Group makes no guarantees regarding the completeness, accuracy, or reliability of any information contained herein. Readers are encouraged to verify facts independently and consult relevant sources for further clarification.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Read the full Futurum Group Disclosure.
Other Insights from Futurum:
DPD Polska's Creatio Rollout: 1,500 Users, Five Weeks, Zero Disruption Qodo Bets on Earned Autonomy to Break the AI Code Review Bottleneck Perplexity's Brain Learns on the Job: Self-Improving Agent Skills

