Summary
Wonderful is the applied AI partner for enterprises in critical industries. It combines a multi-model platform with local deployment teams to help enterprises transform their business with AI. Having raised more than $800 million to date, Wonderful now operates in more than 35 markets worldwide. Voice is one of the most demanding workloads on the platform. For languages that off-the-shelf speech models underserve, Wonderful builds its own multilingual voice models, which have to hold a full conversation at production speed, under strict data residency requirements.
These models run on Together AI’s inference platform, on dedicated NVIDIA GPUs, giving Wonderful full control over scaling, deployment, and model behavior, so its team can spend time on model quality instead of infrastructure.
About Wonderful
Wonderful helps large enterprises become AI native. The Wonderful AI OS is the operating layer that coordinates goals, context, models, agents, workflows, and applications, so AI runs across an organization instead of sitting in isolated point solutions. Every deployment adds to a shared foundation: enterprise context, integrations, and governed execution accumulate over time, so what starts as a single use case becomes a system that gets faster and smarter to build on. Wonderful's forward-deployed engineering pods co-build with customer teams to get a first use case into production, then transfer capability so enterprises increasingly operate the platform themselves.
The challenge: building multilingual voice models in production
Wonderful's model chain has to hold together at conversation speed, every time. Building its own voice models in-house raised that bar in concrete ways:
- Wonderful set out to build its own multilingual voice models: Rather than rely only on off-the-shelf speech models, Wonderful fine-tunes its own models to properly cover languages that general-purpose models underserve. Doing that well in production meant the infrastructure underneath had to get out of the way, so Wonderful's team could spend its time on model quality, not on managing GPUs.
- Deployment had one non-negotiable requirement: strict geolocation and data residency. Every endpoint for this workload has to meet those requirements, running only out of the EU. Together’s multi-region GPU cloud gives Wonderful that control directly, without having to trade compliance against which GPU generations are available to them.
- Scaling needed to move in-house: Every config change or capacity request used to mean a manual, multi-step process: a Together engineer editing configuration by hand, a request filed, a wait for turnaround. Wonderful's pace of growth meant constant iteration, new configurations, A/B tests, that couldn't afford to wait on a manual process. They needed that control to be self-serve.
- Model behavior had to meet Wonderful’s own production bar: Running models that handle live handoffs to downstream systems meant matching Wonderful’s exact production configuration precisely, not just serving a generic version of the model.
- Infrastructure had to be visible immediately: Any infrastructure issue or failed inference needs to be transparent, so Wonderful can monitor their own endpoint health and take recovery action themselves, rather than escalate through a support queue.
"Iteration speed is a competitive advantage for us. With Together, we deploy new model versions on our own timeline, at the performance our customers expect." — Loten Noy, AI Research, Wonderful
The solution: self-serve control, built for a live product
- Full endpoint lifecycle in Wonderful's hands: Together AI's Dedicated Model Inference gives Wonderful direct control over endpoint creation, sizing, scaling policy, and configuration, running on dedicated NVIDIA GPUs in the EU region. Wonderful now handles self-serve scaling, custom weight rollouts, and blue/green and A/B testing on its own timeline, work that used to route through a Together engineer and a ticket queue.
- Direct visibility into endpoint health: A self-serve metrics dashboard gives Wonderful's team request latency, time-to-first-token, throughput, and inference success rate directly, closing the exact gap they had flagged: no more escalating through support every time something needed a look.
- Models configured to meet Wonderful’s production standard: Together matched Wonderful’s exact production configuration for tool-calling behavior, the kind of precision a voice agent needs when it’s handing off to downstream systems mid-conversation.
- Speed built into Wonderful’s voice models: Wonderful’s voice workloads run on GPU clusters located near their own servers, with optimization work across the full stack, including a custom TensorRT engine and quantization, to cut every possible millisecond. The result is among the fastest Whisper serving available across providers.
Results: infrastructure that disappears into the background
For a live voice product, control over the infrastructure underneath has to sit with the team running it. Wonderful's team now manages that infrastructure directly:
- Time-to-first-sentence at 274ms, holding near Wonderful's 300ms target, the point where a text-to-speech engine can start speaking naturally, with zero failed requests across hundreds of calls per test run at production concurrency levels.
- Scaling, rollouts, and testing moved in-house, with zero tickets filed to Together for any of it. What used to be a multi-step request-and-wait process for Wonderful's platform team is now a self-serve action taken by the team that needs it, when they need it.
- More engineering time spent on model quality: With endpoint lifecycle, scaling, and performance tuning handled by the platform, Wonderful’s team spends its energy fine-tuning voice models for the languages its customers need, not managing GPUs underneath them.
For Wonderful, voice is one workload inside a much larger system. The same AI OS that runs this voice function is built to coordinate agents, workflows, and applications across an entire enterprise, with every new deployment adding to a foundation the next one builds on faster. That's the system built to scale, and voice is proof it holds up where the stakes are highest: live, in real time, mid-conversation.


