Smallest.ai

Learning Objectives

• Understand why Smallest.ai builds small speech models instead of large general-purpose ones
• Identify what each of its four first-party models does and where it fits in a voice pipeline
• Evaluate its per-minute pricing against hosted competitors
• Judge which of its performance claims are independently established and which are vendor claims

What Is Smallest.ai?

Smallest.ai is a voice AI company based in Bengaluru, India, founded by Sudarshan Kamath and Akshat Mandloi, both previously engineers at Robert Bosch. In July 2026 it raised a $13 million Series A, led by Seligman Ventures, with Sierra Ventures and 3one4 Capital participating, bringing total funding past $21 million.

The company's argument is narrow and worth stating plainly: in a spoken conversation, latency matters more than raw capability. A model that is marginally smarter but answers a beat too late feels less human than a faster, simpler one. So instead of routing every turn through a large general-purpose model, Smallest.ai trains compact models that listen, reason and speak concurrently, and hands off to a larger third-party model only when a query genuinely needs it.

That design choice is the whole product. It is also the thing to test — the approach lives or dies on whether the small models are good enough for the turns they handle alone.

Key Capabilities

Lightning — text-to-speech

The flagship speech model, advertised at roughly 100 milliseconds of latency across more than 15 languages. The current generation is Lightning v3.1; the company's own documentation marks Lightning v2 and lightning-large as deprecated, so new builds should target v3.1.

Pulse — speech-to-text

Transcription across 38 languages, with speaker separation and emotion detection. Priced far below the text-to-speech side, at roughly one cent per minute.

Hydra — speech-to-speech

A native speech-to-speech model, meaning audio in and audio out without a text round trip in between. This is where the concurrent listen-think-speak design shows up most directly.

Electron — a small language model

A language model under three billion parameters, intended to handle the reasoning inside a voice turn without the cost or latency of a frontier model.

Voice-agents platform

Above the models sits a configuration layer for building and deploying agents — call orchestration, a testing suite, and deployment. This is the surface most buyers actually purchase; the models are what make it cheap.

Pricing

The per-minute range reflects model choice rather than a plan upgrade. Speech-to-text runs around one cent per minute and text-to-speech around nine cents; if you route a turn through a third-party frontier model, that provider's cost passes through on top at cost. The company advertises rates as low as five cents per minute at volume.

For evaluation, the $10 in free credits is the relevant number — it is enough to run real traffic through the stack and measure latency on your own workload rather than trusting the advertised figure.

Smallest.ai vs. Other Voice Platforms

The useful contrast is with Fish Audio, which also sells a hosted platform backed by its own models but publishes downloadable weights under a restrictive research license. Smallest.ai publishes no weights at all — everything is hosted API access. That removes the licensing ambiguity but also removes the self-hosting escape hatch.

Strengths

• Latency as the design goal, not a feature — the concurrent listen-think-speak architecture targets the specific failure that makes voice agents feel robotic
• Transparent per-minute pricing — published rates and $10 in free credits mean you can evaluate it without a sales conversation, which is not true of much of this category
• Full stack in one vendor — text-to-speech, transcription, speech-to-speech and the reasoning model come from one provider, so there is one bill and one latency budget
• Compliance posture aimed at regulated buyers — SOC 2 Type 2, ISO 27001, GDPR and HIPAA claims, with a zero-data-retention option on the enterprise tier
• Real telephony customers — RingCentral and Truecaller are contact-center and telephony businesses, which is a meaningful signal for a voice vendor

Limitations and Considerations

• Early-stage company — a $13 million Series A is small next to the funding behind ElevenLabs or OpenAI, and voice infrastructure is a category where vendors get acquired or repriced
• Frontier claims are self-published — the Electron-versus-GPT-4.1 comparison has no independent evaluation behind it
• No downloadable weights — hosted API only, so there is no self-hosting path and no way to pin a model version outside the vendor's lifecycle
• Deprecation cadence is real — Lightning v2 and lightning-large are already marked deprecated, so plan for migrations
• Third-party model costs pass through — the advertised per-minute rate covers the speech layer; routing turns to a frontier model adds that provider's cost on top

Company Details

Smallest.ai is a private company headquartered in Bengaluru, India. Total funding is above $21 million following the July 2026 Series A, led by Seligman Ventures. No post-money valuation was disclosed. Chief executive Sudarshan Kamath has said the aim is a voice model a caller cannot distinguish from a person.

Key Takeaways

• Smallest.ai sells small, fast speech models rather than frontier capability — the bet is that latency is what makes voice agents feel human
• Four first-party models cover the pipeline: Lightning for speech, Pulse for transcription, Hydra for speech-to-speech and Electron for in-turn reasoning
• Self-serve pricing starts at $10 in free credits with per-minute rates from about nine cents, so it is genuinely evaluable without a sales call
• The claim that Electron beats GPT-4.1 is the company's own and is not independently verified — benchmark it yourself before designing around it
• There are no downloadable weights, so unlike Fish Audio or Voxtral there is no self-hosting fallback if pricing or availability changes