Hopper tunes smaller models to shorten the pauses in voice agents

The pause before a voice agent answers can be more noticeable than the sophistication of its reply. Hopper is targeting that delay with an inference and post-training service built specifically for real-time agents, using open-weight models tuned on a customer's production calls.

The company works across the components of a voice interaction: speech recognition, the language model that decides what to say, and text-to-speech. Its pitch combines running those models efficiently with adapting them to the narrower job a customer needs them to perform.

At launch, Hopper reported an 80-millisecond time to first token for Gemma 4 31B and a 150-millisecond time to first audio for Qwen 3 TTS. It also listed Nemotron 3.5 ASR for speech recognition. These are the company's reported measurements. They do not measure the entire delay a caller experiences across transcription, reasoning, tool use, speech generation and the network.

That distinction is important to the product's appeal. Improving one component can remove part of a conversational pause, but a fast first token does not by itself establish that a complete voice interaction will feel natural. Hopper's service includes deployment and ongoing monitoring, rather than simply access to a model endpoint.

The company describes a workflow that begins with optimizing inference and deploying models in the regions a customer requires. If a base model does not perform well enough, Hopper works with the customer's engineers to design evaluations from its system prompt. Post-training then targets the behavior those evaluations measure.

Some examples are small but audible. A speech model might need to pronounce a company's name correctly, slow down when reading a sequence of letters and numbers, or change its delivery. Language-model work may instead focus on following instructions and choosing the right tools.

Hopper's underlying argument is that a specialized voice agent need not pay for capabilities its task does not use. A lead-qualification system, for example, has a different job from a general-purpose model solving difficult mathematics. The startup expects more voice workloads to shift toward smaller, task-specific models, although the adoption and cost forecasts in its announcement remain predictions.

Founders Pavan Katta and Jashwanth Pedapudi have worked on different parts of inference infrastructure. Katta previously ran voice inference at Salient, while Pedapudi worked on inference kernels for Meta's Quest 3. The company's website says they met in high school more than 13 years ago and later studied at IIT Madras. Hopper is in Y Combinator's Fall 2026 batch.

The launch makes several speed and price comparisons with other providers, but does not establish a universal result across workloads, hardware and configurations. A customer's own evaluations will determine whether a smaller model preserves the behavior its agent needs.

That gives prospective customers a specific tradeoff to examine: how much latency and cost can be removed while the agent still recognizes the caller, follows its instructions and says the right thing. Hopper wants to handle the deployment and tuning work behind that comparison, then keep improving the model against live traffic.