Why small models
Three arguments, one of which we will talk you out of if it does not apply to you.
The latency budget
A voice turn breaks into speech recognition, retrieval, generation, compliance check and synthesis. [Insert your measured breakdown.] Every hop off your network is spent budget you do not get back.
The cost curve
Per-token pricing scales linearly with call volume. Self-hosted inference is a fixed GPU cost with near-zero marginal cost per call. There is a break-even volume, and below it a hosted API is genuinely cheaper. [State your break-even.] We tell you which side of it you are on in the feasibility review.
For Indic-language calls the crossover arrives sooner, because Devanagari tokenises less efficiently than Latin script and per-token pricing charges you for that.
The capability question
Frontier models are better at open-ended reasoning and unconstrained persuasive writing. That is real and we do not argue with it. But most of a regulated call is not open-ended — it is a bounded task with a known good answer, and a fine-tuned small model does bounded tasks at comparable quality for a fraction of the cost and latency.
On code-mixed Indian speech we expect to be better rather than merely comparable, because the training data frontier labs have seen is thin exactly here. That is a claim we intend to prove with numbers, not assert. See methodology.
What we host and what we do not
We fine-tune and serve the language and compliance layer. Speech recognition and synthesis are separate decisions with real quality trade-offs, especially on code-mixed audio. [Fill in your current stack position before launch. Readers will check this, and being precise reads as competence.]