Desert Ant Labs

Desert Ant Labs is building the intelligence layer for every app. Rather than one large model that tries to do everything, the company publishes a family of small, specialized AI models, each of which does one job very well across speech, text, and vision. The models are designed to run on-device — on a phone or in a browser — and are added to any product through one native SDK in just a few lines of code. The company's stated goal is "little brains in every product," giving developers the fastest model for their specific task instead of the overhead of a general-purpose model. The product addresses the cost and dependency that come with cloud-based AI. Because the models run on the device itself, they need no internet connection and involve no per-use or token cost. That matters for two reasons the site calls out directly: developers never have to meter their users, and sensitive data such as personally identifiable information can be filtered on the device instead of being sent away. The company explicitly positions its approach against using one big model for everything, arguing that small models dedicated to a single task deliver the fastest result for that task. Speech, text, and vision each get purpose-built models rather than a single general system. Speech is the deepest area of the model catalog. Voz handles speech recognition, transcribing ten minutes of audio in two seconds on an iPhone. Clear provides speech enhancement for studio sound without a cloud bill. Uhm detects and removes filler words in seconds, which is useful when cleaning up recorded conversations before publishing. Align produces accurate word timestamps for any transcript, even though it is listed more briefly than the others. Ear detects spoken language from 30 seconds of audio, while Tongue identifies language from as little as three words. Together these models cover the path from raw audio to clean, searchable, and well-labelled text, and each one runs on the device, so none of that processing depends on a remote service. On the text side, Gist generates topics and tags for posts and articles, helping content be organised and discovered. Title suggests a title and description for any text, cutting the friction out of publishing. Several models are marked as beta. Schemer performs structured extraction, turning any text into typed JSON, which is directly useful for developers who need machine-readable output from unstructured input. Moderator flags nudity before content is uploaded or displayed, and Toxic is built for hate speech triage, catching hate speech before it posts. Those moderation models are described as running on the device, so content checks happen locally rather than after the fact in the cloud. Vision and media tasks are covered as well. Shapes performs shape recognition, turning a rough sketch into a perfect shape, which suits drawing and design tools where users draw imprecisely and expect clean geometric output. Clips handles clip selection, creating short videos and highlight clips. Emo suggests emoji faster than a user can type, aimed at messaging and social products. Redact filters personally identifiable information on the device, a model the site presents alongside Clear as a way to process sensitive material locally rather than in the cloud. Each of these models is described in one line because each does one narrow job rather than many. The unifying mechanism is a single native SDK. The company describes it as one SDK that drops the models into any product in a few lines of code, which means a developer does not need a different integration for each capability. The models themselves are published on Hugging Face, the SDK is available on GitHub, and documentation is provided separately. Because inference happens on-device, the working method is local execution: the model runs on the user's phone or in their browser rather than calling a remote endpoint. That on-device approach is what removes the need for tokens, logins, and per-use billing from the developer's perspective. The stated benefits follow from that design. Speech enhancement delivers studio sound without a cloud bill; PII redaction keeps sensitive filtering on the device; and transcription is fast enough to process ten minutes of audio in two seconds on an iPhone. Developers can build without metering their users, and the pricing model reinforces this: every model is free up to 100k monthly active devices per platform, with no limit on how often each person runs it. The company frames the outcome simply — build your wildest ideas and best products, and never meter a user. Concrete scenarios follow from the model list. A recording or podcast app can run Voz to transcribe audio and Uhm to strip filler words before publishing, with Align supplying word timestamps for captions or search. A social or community platform can call Moderator before an upload is displayed and Toxic before a comment is posted, checking content on-device. A notes or publishing tool can use Gist for topic tags and Title for suggested titles and descriptions. A drawing app can use Shapes to snap rough sketches into clean shapes, while a messaging app can use Emo for emoji suggestions. A developer pipeline can use Schemer to extract typed JSON from unstructured text, and a privacy-conscious product can run Redact to filter PII locally before data leaves the device. The product is aimed at developers and product teams who are adding speech, text, or vision features to an app and want to avoid cloud costs, per-use token billing, and remote data processing. It is offered as an SDK, with the SDK available on GitHub, documentation on the company's site, and models published on Hugging Face, and it is listed under the topics Artificial Intelligence and SDK. On pricing, every model is free up to 100k monthly active devices per platform, and there is no limit on how often each person runs it. In short, Desert Ant Labs supplies small, task-specific AI models that run on-device and plug into any product through one native SDK. Fast transcription, speech enhancement, on-device redaction, clip selection, and a growing catalog of text and vision models — all free up to 100k monthly active devices per platform — make the pitch simple: the fastest model for the job, with no tokens and no meter.