Desert Ant Labs

Small specialised on-device models for speech, text, and vision, one per task

Editorially reviewedChecked Sep 2026How we review

ai-infrastructure#on-device models#edge ai#speech recognition#privacy

Free plan Self-hosted Mobile app

Toolglade’s take

This is infrastructure for app developers rather than a tool you sit down and use, and on those terms it is well judged. Small per-task models running on-device remove the three costs that make cloud inference painful in consumer apps: latency, per-call billing, and the privacy conversation. The 100,000 monthly active device threshold is generous enough that most apps will never pay. Two things to be precise about. The licence is source-available, explicitly not open source, and attribution is required, so read it before shipping. And commercial pricing above the threshold is not published, which makes the cost of success unknowable in advance.

About Desert Ant Labs

Desert Ant Labs publishes small task-specific models that run entirely on-device across speech, text, and vision, delivered via Swift, Kotlin, and JavaScript SDKs and free below 100,000 monthly active devices.

Desert Ant Labs takes the opposite bet to frontier model vendors: rather than one large model doing everything through a cloud API, it publishes many small models that each do one job and run on the user device. The catalogue includes Voz for speech recognition, Align, Clear, Uhm, Ear, and Clips for audio, Redact, Gist, Title, Tongue, and Emo for text, and Shapes for vision, with further models in beta. The practical consequence is that there is no cloud API and no per-call cost. You embed the model via a Swift, Kotlin, or JavaScript SDK and it executes locally, which removes latency, works offline, and means user audio or text never leaves the handset. Weights are published on Hugging Face. Licensing is source-available rather than open source. The Desert Ant Labs Source-Available Licence 1.0 permits free use below 100,000 monthly active devices per platform and per model, with no limit on how often each person runs it and no per-call fee; research and teaching are exempt from the threshold. Above it, a commercial licence is required. The company is a Dutch entity, Desert Ant Labs B.V.

Weighing your options?See how Desert Ant Labs compares to the alternatives.

TL;DR

Desert Ant Labs publishes small, one-task-each models that run on-device, free below 100,000 monthly active devices.

Company overview

Desert Ant Labs B.V. is a Dutch company publishing small specialised models designed to run on end-user devices rather than in the cloud.

It was featured on Product Hunt in September 2026 and distributes model weights publicly on Hugging Face under a source-available licence.

Product features

The catalogue spans audio models including Voz, Align, Clear, Uhm, Ear, and Clips, text models including Redact, Gist, Title, Tongue, and Emo, and vision models including Shapes, with further beta models.

Delivery is via Swift, Kotlin, and JavaScript SDKs with no cloud API and no per-call cost. Use is free below 100,000 monthly active devices per platform and model.

Target market

Mobile and web application developers who need AI capability without cloud latency, per-call billing, or sending user data off-device.

Buyer personas

End users

Developers embedding models in applications.

Buyers

Product and engineering leads at app companies.

Key influencers

Edge AI and on-device machine learning communities.

Ideal customer profile

A consumer app team that needs speech or text processing at scale without per-call inference costs or a privacy review.

Funding & performance

Not stated; the entity is Desert Ant Labs B.V., governed by Netherlands law.

Pros & cons

Pros

Runs fully on-device with no cloud API or per-call cost
Free below 100,000 monthly active devices per platform and model
No limit on how often each user runs a model
Weights published publicly on Hugging Face
Swift, Kotlin and JavaScript SDKs
Research and teaching exempt from the device threshold
Works offline with no latency from network calls

Cons

Source-available licence, explicitly not open source
Commercial pricing above the threshold is not published
Attribution is required by the licence
Each model covers one narrow task
Developer audience only, not an end-user tool

Pricing plans

Free tier

Free below 100,000 monthly active devices per platform per model
No per-call cost
No limit on runs per person
Research and teaching exempt

Commercial

Contact vendor

Required above 100,000 monthly active devices
Pricing not published

Key features

Mobile app

Self-hosted

Multi-language

Integrations

Swift, Kotlin, JavaScript, Hugging Face, GitHub

Input types

audio, text, image

Output types

text

Best For

on-device speech recognition, offline ai in mobile apps, privacy-preserving inference

Compare key features

View all alternatives →

Feature

Desert Ant Labs

Ollama

OpenRouter

Pricing

Freemium

Free plan

Yes

Free trial

API

Yes

Self-hosted

Yes

Team support

Yes

Frequently asked questions

Do these models run in the cloud?+

No. They run entirely on the device via Swift, Kotlin, or JavaScript SDKs, with no cloud API.

When do I have to pay?+

Above 100,000 monthly active devices per platform and per model, which requires a commercial licence from the vendor.

Is it open source?+

No. It uses the Desert Ant Labs Source-Available Licence 1.0, which the vendor explicitly describes as source-available rather than open source.

Which tasks are covered?+

Speech recognition, alignment, noise clearing, filler removal, clipping, redaction, summarisation, titling, translation, emotion, and vision, with more models in beta.

Where can I get the weights?+

Model weights are published publicly on Hugging Face.