Desert Ant Labs

Desert Ant Labs builds small, specialized AI models for speech, text, and vision, and delivers them through one native SDK that developers can drop into any product in a few lines of code. Instead of relying on a single large model to handle every task, the library offers a family of focused models where each one does a single job very well — from speech recognition and speech enhancement to PII redaction, content moderation, and structured extraction. The models run on the user's phone or in the browser, with no internet connection required. Most AI-powered product features depend on a cloud service. Sending audio, text, or images to a remote endpoint means paying per use, requiring a network connection, and moving user data off the device. Desert Ant Labs positions itself against that model: its models run on-device, so there is no internet requirement, no token cost, and no need to meter a user. The company describes its work as building "the intelligence layer for every app" — a set of small models that each do one job very well, with one native SDK that drops them into any product. The stated aim is that builders can pursue their wildest ideas and best products and never meter a user. Speech and audio are the deepest part of the library. Voz handles speech recognition and can transcribe ten minutes of audio in about two seconds on an iPhone. Clear is a speech enhancement model that produces studio sound without a cloud bill. Align generates accurate word timestamps for any transcript, which is the groundwork for captioning, karaoke-style highlighting, and clips that start and end on the right words. Uhm detects filler words so they can be found and removed in seconds. Ear performs spoken language detection from just 30 seconds of audio, while Tongue identifies a language from as few as three words. Together these models cover a pipeline from raw audio to a cleaned, timestamped, language-tagged transcript. On the text side, Redact filters personally identifiable information on the device, so sensitive data can be caught before it leaves the app or is stored. Schemer, currently in beta, performs structured extraction and turns any text into typed JSON. Gist generates topics and tags for posts and articles, and Title suggests a title and description for any text. Emo suggests emoji faster than a person can type them. For safety, Moderator (beta) flags nudity before content is uploaded or displayed, and Toxic (beta) triages hate speech to catch it before it posts. Each of these is a narrow, task-specific model rather than an open-ended assistant, which is the core idea behind the library: small models that each nail one job. On the media and vision side, Clips handles clip selection and creates short videos and highlight clips from longer footage. Shapes is a shape recognition model that turns a rough sketch into a perfect shape, useful for drawing and diagramming interfaces where a user's hand-drawn input needs to be cleaned up. Alongside Moderator's role in flagging nudity before upload or display, these models extend the platform from language tasks into media and visual input. Every model is delivered through one native SDK. Developers add a model to their app in a few lines of code rather than integrating a separate service for each capability, and there is a try-it-for-free path with no tokens and no logins. Because inference happens on-device, the model runs against local input on the phone or in the browser. The catalogue spans speech, text, and vision, and the beta models for content moderation and structured extraction show the library continuing to expand. The models are also published on Hugging Face, so developers can evaluate them directly. The clearest benefit stated is cost: with no per-use charge and no token metering, a product can run AI features without a bill that scales with usage, and the free tier covers up to 100,000 monthly active devices per platform with no limit on how often each person runs a model. The second benefit is privacy and control, since inference happens on-device and input does not need to be sent to a cloud service. The third is speed and reliability: running locally means results do not wait on a network round trip, and the features keep working without a connection — Voz's ability to transcribe ten minutes of audio in roughly two seconds on an iPhone is presented as an example of that on-device performance. Concretely, the models map to common product workflows. A recording, meeting, or podcast app can use Voz for transcription, Align for word timestamps, Uhm to find and remove filler words, and Clear to enhance the audio to studio quality. A video tool can use Clips to create shorts and highlights from longer footage, with Align providing accurate word timestamps to choose cut points. A publishing or messaging app can use Gist to generate topics and tags for posts and articles, Title to suggest a title and description for any text, and Emo to suggest emoji faster than a user can type. A platform handling user-generated content can run Moderator to flag nudity before upload or display, Toxic to catch hate speech before it posts, and Redact to filter PII on the device. A sketching tool can use Shapes to turn a rough sketch into a perfect shape, and a data workflow can use Schemer to extract typed JSON from any text. Multilingual apps can detect the spoken language with Ear or identify a language from three words with Tongue. Desert Ant Labs targets developers and product teams adding AI capabilities to their own applications — mobile and web products that need speech, text, or vision features without a cloud dependency or per-use pricing. The models are free up to 100,000 monthly active devices per platform, with no limit on how often each person runs a model, and there is no login or token requirement to try them. Supporting resources include the SDK on GitHub, documentation on the Desert Ant Labs site, and the models published on Hugging Face, which gives developers several ways to review and integrate the technology before shipping. Desert Ant Labs is best understood as an intelligence layer for apps that want AI features without the usual cloud tax. By splitting capability into small, task-specific models for speech, text, and vision and shipping them through a single SDK that runs on-device, it lets builders add transcription, speech enhancement, redaction, moderation, tagging, clip selection, and structured extraction in a few lines of code — free up to 100,000 monthly active devices per platform.