Smallest AI

What Smallest AI is

Smallest AI is a voice AI company that sells 2 things at once: its own speech and language models through an API, and a voice agent product built on top of them. The home page calls it a "Realtime voice AI suite built for enterprise scale," with Agents, Models, Resources and Pricing as its main sections. It is written first for developers and for teams building real-time voice agents, and second for businesses that want an agent without assembling the stack themselves. That puts it at the code-friendly end of this category: a platform with a point of view on model size rather than a drag-and-drop canvas. The footer gives a San Francisco office address on California Street, and lists the industries it addresses: healthcare, real estate, travel and hospitality, automobile, banking and e-commerce.

What you build and how

The agent side is introduced with a short line: "Voice Agents. Agents that keep the conversation going. Any language. Any channel." The hero demo shows 2 example agents, a Support Agent and a Call agent. Further down, the home page describes "The agentic platform for every use case" and says you "Configure your agent, pick your voice, set your languages, and go live all from a single interface built for production." An Agents Playground is shown beside that text. The company's blog adds the product name: voice agents come through Atoms, which it describes as STT, TTS and orchestration in one place. In practice the build loop is configuration rather than diagramming: choose the voice, set the languages the agent should handle, attach it to a channel and test it in the playground before it goes live.

The models are where most of the detail sits. The home page lists these:

• Lightning v3.1 Pro, a text-to-speech model described as "Blazing-fast text-to-speech for real-time voice agents." The hero says sub 100ms latency across 70+ languages, while the model card further down says 100 ms latency across 15+ languages, so the home page itself gives 2 language counts for the same model.
• Pulse, real-time speech to text, across 38+ languages with 64ms latency. Its model card promises "The most accurate transcription in 38 languages with emotion and speaker detection."
• Electron, which the home page describes as a sub-3B small language model.
• Hydra, labeled Beta, a full-duplex speech-to-speech model that "listens, reasons, and responds in real time." The company calls it one of the first native speech-to-speech models built for production.

Each model card carries a Get API button, and the hero lets a visitor type text, pick a voice and generate speech on the spot, so trying a voice takes seconds.

For a developer, the home page shows a JavaScript example that posts text to the Lightning v3.1 endpoint with an API key, a voice id, a sample rate and an output format, and writes the result to a WAV file. The section around it reads "Start with code. Scale without limits." The company's blog says the API supports streaming output, so audio starts playing before generation is complete, and that there are Python and Node.js SDKs and a REST interface. It also says instant voice cloning works from a few seconds of audio and that Lightning supports 30+ languages and regional accents, a third language figure beside the 2 on the home page. A banner on the blog advertises Lightning V3 with 200+ voices and voice cloning, and the footer links a Voice Library, an Integrations page, separate documentation for agents and for models, and an On Prem entry under Agents. The footer also names 2 further products, LazyVoice and CapKit, and a Startup Grants program, without describing them.

The research section explains the thesis behind the products. The company says it believes AI will evolve like human intelligence: not one massive model that knows everything but many small ones that each know what matters. It argues that small models, with access to external memory and tools, can be fast, efficient and deployable anywhere, from edge devices to real-time applications, and that voice should be treated as a native modality rather than text converted to speech. You do not need to buy the philosophy to use the API, but it explains why the product leads with latency.

On the chatbot side, nothing on these pages describes a text chatbot builder or a website chat widget. Smallest earns its place in the chatbot-builders category through the conversational agent product: agents that hold a spoken conversation, with the promise that they work in any language and over any channel. Read it as a voice-first agent platform and model provider.

What the pages say about pricing and trust

The company's blog lists 3 plans. Free gives 30 minutes of text-to-speech per month. Basic is $5 a month and covers 3 hours of TTS plus API access and 1 voice clone. Premium is $29 a month and covers 24 hours of TTS and 2 voice clones. The same post says the free tier is "enough to build and test a real application before paying anything." Those plan details describe text-to-speech usage rather than Atoms voice agent usage, for which the pages show no price.

The same post claims Lightning runs at sub-100ms time to first byte, and it is open about where the product is weaker: studio-style narration tools such as scripts, timecodes and multi-voice projects are limited, and the post says the API, rather than the web UI, gives the better experience.

On security, the home page says Smallest is "Trusted by enterprises for secure, compliant, production-ready performance," and a section headed "Guarding your data with enterprise security" shows ISO 27001, SOC 2 Type 2, GDPR Compliant and HIPAA Compliant badges with a View compliance link. The footer repeats "We are SOC 2, GDPR, and HIPAA, Compliant," links a TCPA compliance policy and a data processing page, and shows a system status indicator. A self-serve Sign Up path sits beside Contact Sales, so a team can try the free tier first and talk to sales about larger deployments.

Who it suits and who it does not

Smallest suits an engineering team building a voice agent or IVR that cares about response time and wants to choose its own components. A startup can start on the free 30 minutes, and a product team can use the speech-to-text and text-to-speech APIs inside an existing stack. A team that wants one vendor for STT, TTS and the agent layer can use Atoms, and a buyer in healthcare or another regulated field will find the compliance badges stated on the home page.

It suits less well a team that wants a visual flow builder with templates, a text-based website chatbot, or long-form narration work, which the company's own post says is not its strength. The language counts differ between the home page and the blog, and the pages publish no Atoms pricing, so a buyer comparing it with a no-code agent builder should plan a test in the playground and against the API before committing.