# Desert Ant Labs: 12 On-Device AI Models for iOS, Android, and the Web Desert Ant Labs is building a different kind of AI stack: **small, single-purpose models that run entirely on-device**. There is no server to call, no API key to manage, no token bill, and no network connection required once the model is available locally. On Apple platforms, the models target the **Apple Neural Engine through Core ML**. Android support comes through LiteRT, while SDKs are available for Swift, Kotlin, and JavaScript, with WebAssembly support for the web. Model weights are published on Hugging Face, and there is also a CLI for macOS and Linux. The company describes this approach as the **“cerebellum layer” of AI**. The idea is that fast, repetitive, reflex-like tasks should happen locally, leaving larger cloud models to handle more complex reasoning. Instead of sending every small decision to a general-purpose model, an app can offload narrow tasks to lightweight models running directly on the device. There are currently **12 publicly available models, with five more in beta**. Usage is free for up to 100,000 monthly active devices per platform, after which a commercial license is required. One important detail is the license itself: the models use `desert-ant-labs-source-available-1.0`, not MIT or Apache, so **the terms are worth reviewing carefully before using them in a commercial product**. The GitHub organization is available at [github.com/Desert-Ant-Labs](https://github.com/Desert-Ant-Labs), with `desert-ant-core` serving as the main SDK repository. Current platform requirements include **iOS 18+, macOS 15+, tvOS 18+, visionOS 2+, Swift 6.2+, and Xcode 26**. Desert Ant Labs also pins each SDK release to a specific model revision. That means **the model cannot silently change underneath a running application**, and every downloaded model is verified before use. For teams building on-device AI features, that kind of versioning makes behavior easier to reproduce and control. ## The 12 models ### Voz: speech recognition Voz transcribes spoken audio to text with word-level timestamps. The claim is 10 minutes of audio in 2 seconds on an iPhone, roughly 4.7x faster than Whisper. It supports 25 languages. A 7-minute recording in testing completed transcription in 2.04 seconds with accurate word-level sync confirmed by playback. ![The interface of the custom Voz application after transcribing a 7-minute, 4-second audio file. The metrics clearly show the task was completed in just 2.04 seconds.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/6914fc40-a2b6-4c7c-0084-b01f1b523e00/lg2x =1920x1080) The timestamps are precise enough to support karaoke-style caption display and transcript-driven video editing. ### Clear: speech enhancement Clear removes background noise, reverb, and mic bumps, and normalizes loudness toward a close-miked studio sound. It handles HVAC, keyboard clicks, room hum, laptop fans, and coffee shop ambient without degrading consonants. Works on 48kHz mono or stereo audio from meeting recorders, Bluetooth microphones, phone mics, and laptop microphones. ![The "Clear" model's interface in the testing app. It features two tabs, "Original" and "Clear," allowing for an instant A/B comparison of the audio waveforms before and after processing.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/945f5888-35ae-4fb4-8c7e-a124cf431600/md2x =1920x1080) In testing against a loud kitchen fan, noise was reduced noticeably and voice presence improved significantly, though not completely. Estimated at roughly 70% effective in that extreme environment. For moderate noise (coffee shop, open office), results are stronger. ### Clips: intelligent clip selection Clips analyzes a long recording, transcribes it, and identifies and ranks the best segments for use as short-form clips or highlights. The benchmark on the product page: processing 100,000 thirty-minute videos using 470x less energy than cloud-based Claude Sonnet on the same task. The companion desktop app Clipper (available at `Desert-Ant-Labs/demo-clipper`) provides a drag-and-drop interface for the full workflow. ![The Clipper desktop application with a video loaded, showing the list of automatically generated "top clips" on the left and a detailed view of the selected clip, including its transcript and title, on the right.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/c29c9804-86b5-4677-a67a-e04a31450e00/lg1x =1920x1080) In testing with a 1-hour 47-minute podcast, Clipper generated six ranked clips in about 3 minutes. Each came with a generated title and summary (using the Title model), exact in-out timestamps, and one-click export. The transcript-driven editing lets you cut video by editing text. ### Redact: on-device PII removal Redact detects and removes personally identifiable information from text before it leaves the device. Supported entity types include given names, surnames, phone numbers, email addresses, URLs, physical addresses, IP addresses, passport numbers, credit card numbers, IBAN codes, social security numbers, and Bitcoin addresses. In testing: "Email Anna at anna@example.hu or call 555-0100" becomes "Email [GIVEN_NAME_1] at [EMAIL_1] or call [PHONE_1]". Credit card numbers, SSNs, and addresses redacted correctly. One miss: "John lives at 123 Main Street" failed to redact the street address. Strong overall, but requiring exact redaction coverage will need testing against your specific data types. ### Filler: filler word removal Filler detects and removes spoken filler words and phrases from transcripts: "um", "uh", "like", "you know", "basically", "I mean", and similar. Tested against a transcript with 37 instances of "um" and "uh": 36 removed correctly, one missed. The model also catches less obvious fillers like "kind of" and "sort of" used as hedges rather than as meaningful phrases, though this is context-dependent. ![The "Filler" model's interactive demo. It shows input text riddled with filler words like "um," "uh," and "kind of," and the cleaned output text below with those words removed.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/4f4dd0c3-b00e-430c-393d-23ba5acf3600/public =1920x1080) This model composes naturally with Voz: transcribe with Voz, clean fillers with Filler, then pass the result to a summarization model. ### Gist: topic classification Gist classifies a piece of text into one or more predefined topic categories with confidence scores. Testing results were inconsistent. A study about music's effect on studying was classified as "Arts & Culture" (plausible but not the best fit). A headline about sleep research was correctly classified as "Science & Technology." A headline about a space mission arriving at Mercury was classified as "Careers & Productivity" (incorrect). ![The "Gist" model's on-page demo. It shows the input text field and the corresponding topics generated below with confidence scores represented by progress bars.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/ba101095-e6a8-4e1d-df4e-ea2d3e221400/md1x =1920x1080) The failures suggest keyword matching rather than semantic understanding. Mercury in a space context appears to be associated with "careers" rather than "science." Not production-ready for general topic classification without testing against your specific category set. ### Align: word timestamp alignment Align generates accurate word-level timestamps for any transcript by analyzing the audio to find each word's precise start and end time, rather than estimating from model output. The model is 0.7 MB. It's a backend utility rather than a standalone tool: it powers the precise video cuts in Clipper and enables frame-accurate caption display. Five times more accurate than standard transcribers on the benchmark, though this is difficult to verify directly without a ground-truth dataset. ### Shapes: on-device shape recognition Shapes recognizes a hand-drawn stroke and converts it instantly into a perfect geometric shape (rectangle, ellipse, triangle, star, line) in under 10 milliseconds. Single-stroke shapes (square, triangle, star drawn in one motion, oval drawn continuously) snap correctly and instantly. Multi-stroke shapes (an ellipse drawn with two separate arcs) are not recognized as a single shape. A reliable tool for single-stroke drawing applications; multi-stroke compound shapes are outside its current capability. ### Title: title and description generation Title generates a short factual title and a one-to-two sentence description for any given passage of text. This model powers the clip titles and summaries in Clipper. Useful as a composable component in document processing pipelines where you need short-form summaries of text chunks. ### Tongue: written language identification Tongue identifies the language of a piece of text in tens of microseconds. Supports 84 languages. The 2 MB model ships inside each SDK package, so it requires no download at first use. French: correct at 99% confidence. Latvian phrase: correct at 99.6% confidence. "Mitsubishi" as a standalone word: guessed Polish at 56% confidence, incorrect. Single proper nouns without linguistic context are genuinely difficult for any language identifier. For short sentences with enough content words, performance is reliable. ### Ear: spoken language identification Ear identifies the spoken language from a short audio sample before transcription begins. Supports 99 languages. English: correct at 100% confidence. French: correct at 96% confidence. Latvian: returned Turkish at 29% confidence, incorrect. Processing time in testing was 5-10 seconds, noticeably slower than the other models. Inconsistent across languages and slower than ideal for a pre-transcription gate. More training on lower-resource languages would improve reliability. ![The "Ear" app's interface showing a successful identification of English with a 100% confidence score. The metrics also display that the identification took about 10.5 seconds.](https://imagedelivery.net/xZXo0QFi-1_4Zimer-T0XQ/0b40bdbf-93cc-4524-c8db-c944f8624300/md2x =1920x1080) ## Five models in beta Five additional models are listed as in beta: Eye (image tagging and aesthetics scoring), Face (on-device face matching across a photo library or video), Moderator (NSFW image detection trained on licensed and synthetic data), Schemer (structured JSON extraction from unstructured text), Toxic (hate speech detection for European languages), and Who (speaker labeling with per-person timestamps). The GitHub org lists `toxic-en` and `redact` as recently updated, suggesting active development. ## The CLI A CLI (`da`) runs the same models from the terminal on macOS (Apple Silicon) and Linux: ```command brew install desert-ant-labs/tap/desertant ``` ```command da redact "Email Anna at anna@example.hu or call 555-0100" ``` The CLI supports JSON output, model chaining, and is documented for use with coding agents. Useful for scripting document processing pipelines or testing model output without building an app. ## When to use these models The strongest argument for Desert Ant Labs is **how well the models can be chained together**. Several of them become much more useful when combined into a local workflow rather than used on their own. For example, **Voz + Filler + Redact** can turn speech into a cleaned, PII-free transcript without sending anything to the cloud. **Voz + Align + Clips + Title** can produce timestamped highlights and generate titles for each segment. The reliability of individual models varies, but the broader pattern is compelling: **small, specialized models handling one task each, all on-device**. The pricing also makes experimentation practical. The free tier covers up to **100,000 monthly active devices per platform**, which is enough for meaningful production use in many apps. Once you cross that threshold, a commercial license is required. The main caveat is licensing. Desert Ant Labs uses a **source-available license rather than MIT or Apache**, so you should review the terms carefully before shipping or distributing a commercial product. Each model is documented separately at [huggingface.co/desert-ant-labs](https://huggingface.co/desert-ant-labs), with installation and usage instructions for every supported platform. That makes it relatively easy to evaluate individual models before deciding how to combine them into a larger on-device pipeline.