Desert Ant Labs is a Netherlands-based company (Desert Ant Labs B.V.) that publishes small, task-specific AI models with SDKs for Swift, Kotlin and JavaScript. The models run on the user's phone, Mac, Windows PC, browser tab or Node server through Core ML, LiteRT and WebAssembly, so text, audio and images stay on the device. As of September 2026 twelve models have SDKs, including Redact for PII redaction, Tongue for language detection, Voz for speech recognition and Clear for speech enhancement, and six more are in closed beta.

What is a tiny AI model?

A tiny model is a neural network trained for one narrow task and compressed to run on ordinary hardware, often in a few megabytes. It cannot chat or write like ChatGPT, but on a single job such as detecting a language, masking personal data or transcribing speech it can match much larger models. Desert Ant's 2 MB Tongue model scored 0.933 accuracy on three-word FLORES-200 snippets against 0.887 for the 293 MB lingua library.

How much does Desert Ant cost?

The models are free below 100,000 monthly active devices, counted separately for each model on each platform, and inference per device is unlimited. Above the threshold you need a commercial licence from Desert Ant Labs, priced on request. Non-commercial research and teaching are free at any scale. Free use requires a visible 'Powered by Desert Ant Labs' credit, for example on an about or settings page.

Is Desert Ant open source?

No. It is source-available under the Desert Ant Labs Source-Available License 1.0, dated 3 July 2026. You can read, modify and embed the code and weights in your own product, but you cannot redistribute the models on their own, offer them as a hosted service, or use their outputs to train a competing model. The licence is governed by Dutch law.

Does any data leave the device with Desert Ant?

Your inputs and outputs do not. The model weights download once from Hugging Face, or from a folder you host yourself for offline and airgapped installs. The SDK also sends Desert Ant an anonymous device identifier and call count to meter the free tier, with no user content in it. Desert Ant is the data controller for that telemetry, and the licence asks you to mention it in your privacy notice where the law requires.

Can Desert Ant replace Amazon Comprehend or Whisper?

For some jobs. Redact covers 27 languages where Amazon Comprehend's PII detection supports English and Spanish, and it scored 99.6% precision on Desert Ant's structured test set, but it is behind Comprehend on English names. Voz reached a 7.40% word error rate against 7.00% for Whisper large-v3-turbo at under a third of the size, but it runs only on Apple devices and covers 25 languages. Test either on your own data before switching.

Which platforms does Desert Ant support?

iOS 18, macOS 15, tvOS 18 and visionOS 2 or later through Core ML; Android API 24 or later on arm64 and x86_64 through LiteRT; Windows x64; any browser with WebAssembly; and Node on Linux x64, Linux arm64 and Apple silicon. Some models are Apple-only, including Voz, Title, Uhm and Align. AWS Lambda on arm64 needs a preloaded shim library that ships with the package.

Is on-device AI compliant with UK GDPR and the EU AI Act?

It makes compliance easier without doing it for you. Processing personal data on the device supports the UK GDPR data minimisation principle and avoids adding a processor or an international transfer for that step. You remain the controller for what your application does. Under the EU AI Act, Desert Ant treats its models as narrow-purpose components rather than general-purpose AI models, and you are the provider of the AI system you build, including Article 50 transparency where it interacts with people.

What can't tiny models do?

They do not reason, write or answer open questions: none of Desert Ant's models drafts a customer reply or summarises a contract. Several have published weak spots. Voz is Apple-only and produces confident nonsense on languages it does not cover, Ear cannot reliably tell Norwegian, Swedish and Danish apart, Tongue cannot separate Malay from Indonesian, and Clips and Title have no published quality figures yet. The usual design puts a tiny model in front of a larger LLM as a filter.

How does Cloud First integrate tiny models into an existing setup?

We start with one workflow, usually the support inbox, recorded calls or consultations, or staff use of AI tools, and test the relevant Desert Ant models on a sample of your own data. Then we fit them into what you already run: a Node webhook service for Zendesk, Freshdesk or HubSpot on Google Cloud Run or AWS Lambda, a managed browser extension, a Mac transcription station feeding SharePoint or Google Drive, or the SDKs inside your own app. We also handle the attribution credit, the privacy notice wording and tracking device counts against the free tier.

Tiny Models, Real Jobs: On-Device AI for UK and EU Small Businesses

Security and privacy are the top reasons UK small businesses hold back on AI (44% of owners, Simply Business, August 2026). Desert Ant Labs, a Dutch company, publishes task-specific AI models with Swift, Kotlin and JavaScript SDKs that run on the device through Core ML, LiteRT and WebAssembly: Redact masks personal data in 27 languages at 11.6 MB (88.8% recall and 99.6% precision on Desert Ant's benchmark, against 60.2% and 93.5% for a 3 GB OpenAI privacy filter), Tongue detects 84 languages from a 2 MB model, Voz transcribes 25 languages at a 7.40% word error rate against 7.00% for Whisper large-v3-turbo, Clear cleans speech 302 times faster than real time, and Gist tags content in 101 languages. The article covers six UK and EU small business scenarios (a multilingual e-commerce support inbox, redaction before staff use AI tools in accountancy and law practices, on-Mac transcription for clinics and recruiters, voice notes for field services, archive tagging and comment triage under the Online Safety Act, and webinar clips for agencies), costs against Amazon Comprehend ($0.0001 per 100 characters, PII in English and Spanish only) and Amazon Transcribe ($0.006 per minute batch), a comparison table, the published limits, the source-available licence (free below 100,000 monthly active devices per model per platform, attribution, anonymous usage telemetry, EU AI Act roles), integration patterns for helpdesks, browser extensions, Mac transcription stations, mobile apps and airgapped installs, and a live demo in which the 2 MB Tongue model routes a European support inbox in the browser.

Frequently Asked Questions

What is Desert Ant?: Desert Ant Labs is a Netherlands-based company (Desert Ant Labs B.V.) that publishes small, task-specific AI models with SDKs for Swift, Kotlin and JavaScript. The models run on the user's phone, Mac, Windows PC, browser tab or Node server through Core ML, LiteRT and WebAssembly, so text, audio and images stay on the device. As of September 2026 twelve models have SDKs, including Redact for PII redaction, Tongue for language detection, Voz for speech recognition and Clear for speech enhancement, and six more are in closed beta.
What is a tiny AI model?: A tiny model is a neural network trained for one narrow task and compressed to run on ordinary hardware, often in a few megabytes. It cannot chat or write like ChatGPT, but on a single job such as detecting a language, masking personal data or transcribing speech it can match much larger models. Desert Ant's 2 MB Tongue model scored 0.933 accuracy on three-word FLORES-200 snippets against 0.887 for the 293 MB lingua library.
How much does Desert Ant cost?: The models are free below 100,000 monthly active devices, counted separately for each model on each platform, and inference per device is unlimited. Above the threshold you need a commercial licence from Desert Ant Labs, priced on request. Non-commercial research and teaching are free at any scale. Free use requires a visible 'Powered by Desert Ant Labs' credit, for example on an about or settings page.
Is Desert Ant open source?: No. It is source-available under the Desert Ant Labs Source-Available License 1.0, dated 3 July 2026. You can read, modify and embed the code and weights in your own product, but you cannot redistribute the models on their own, offer them as a hosted service, or use their outputs to train a competing model. The licence is governed by Dutch law.
Does any data leave the device with Desert Ant?: Your inputs and outputs do not. The model weights download once from Hugging Face, or from a folder you host yourself for offline and airgapped installs. The SDK also sends Desert Ant an anonymous device identifier and call count to meter the free tier, with no user content in it. Desert Ant is the data controller for that telemetry, and the licence asks you to mention it in your privacy notice where the law requires.
Can Desert Ant replace Amazon Comprehend or Whisper?: For some jobs. Redact covers 27 languages where Amazon Comprehend's PII detection supports English and Spanish, and it scored 99.6% precision on Desert Ant's structured test set, but it is behind Comprehend on English names. Voz reached a 7.40% word error rate against 7.00% for Whisper large-v3-turbo at under a third of the size, but it runs only on Apple devices and covers 25 languages. Test either on your own data before switching.
Which platforms does Desert Ant support?: iOS 18, macOS 15, tvOS 18 and visionOS 2 or later through Core ML; Android API 24 or later on arm64 and x86_64 through LiteRT; Windows x64; any browser with WebAssembly; and Node on Linux x64, Linux arm64 and Apple silicon. Some models are Apple-only, including Voz, Title, Uhm and Align. AWS Lambda on arm64 needs a preloaded shim library that ships with the package.
Is on-device AI compliant with UK GDPR and the EU AI Act?: It makes compliance easier without doing it for you. Processing personal data on the device supports the UK GDPR data minimisation principle and avoids adding a processor or an international transfer for that step. You remain the controller for what your application does. Under the EU AI Act, Desert Ant treats its models as narrow-purpose components rather than general-purpose AI models, and you are the provider of the AI system you build, including Article 50 transparency where it interacts with people.
What can't tiny models do?: They do not reason, write or answer open questions: none of Desert Ant's models drafts a customer reply or summarises a contract. Several have published weak spots. Voz is Apple-only and produces confident nonsense on languages it does not cover, Ear cannot reliably tell Norwegian, Swedish and Danish apart, Tongue cannot separate Malay from Indonesian, and Clips and Title have no published quality figures yet. The usual design puts a tiny model in front of a larger LLM as a filter.
How does Cloud First integrate tiny models into an existing setup?: We start with one workflow, usually the support inbox, recorded calls or consultations, or staff use of AI tools, and test the relevant Desert Ant models on a sample of your own data. Then we fit them into what you already run: a Node webhook service for Zendesk, Freshdesk or HubSpot on Google Cloud Run or AWS Lambda, a managed browser extension, a Mac transcription station feeding SharePoint or Google Drive, or the SDKs inside your own app. We also handle the attribution credit, the privacy notice wording and tracking device counts against the free tier.

Our Services

Contact Cloud First Consulting

Email: info@cloudfirstconsulting.com

Location: London, United Kingdom

Hours: Monday-Friday, 9:00 AM - 6:00 PM GMT

Book a Free 30-min Discovery Call