A New AI Lab Shipped 18 Models. Its Headliner Is NVIDIA's.

TL;DR

On September 8 Desert Ant Labs launched out of Amsterdam with 18 on-device models, one SDK for Swift, Kotlin and JavaScript, and a free tier that runs to 100,000 monthly active devices per model per platform. It hit the Hacker News front page at roughly 390 points. Then somebody opened the third-party notices files.

Voz, the transcription model in every demo, is NVIDIA's Parakeet TDT 0.6B v3 compiled for the Apple Neural Engine. Its own notices file states that the weight values are unchanged. Clear is a DeepFilterNet 3 fine-tune, Ear is the language head carved out of whisper-tiny, and Clips is XLM-RoBERTa base wearing four new heads. None of that makes this a bad product. It makes it a different product than the launch post described, and the best thing in the drop is a model nobody talked about.

What actually shipped

Eighteen models, 12 stable and six in beta, split across audio, text and vision. Seventeen of them are public repos in the desert-ant-labs Hugging Face org, and the earliest was pushed on May 12, so this launch was a curtain going up on work that had been sitting in the open for four months.

The delivery mechanism is the point. One SDK repo covers iOS, macOS, Android, browser and Node, with Core ML on Apple silicon, LiteRT on Android, and WebAssembly in the browser running the same weights. There is a CLI if you want to try a model on your Mac before you write a line of app code. Installing Redact is one SwiftPM line, one Gradle coordinate, or one npm i.

And the speed is real. Over 30 continuous minutes of audio on an M3 Ultra, Voz runs at 319x realtime. Apple's own SpeechAnalyzer manages 78x on the same job and Whisper large-v3-turbo 50x.

Then somebody read the notices

Within an hour of the post landing, a commenter had a list: Voz is Parakeet 0.6B v3, Clear is DeepFilterNet 3, Ear is the language predictor from whisper-tiny. Founder Paul Veugen replied in the thread and confirmed it without flinching, saying Voz is an ANE-optimized Parakeet with their own inference, and that the next-generation Voz is trained from scratch and should be twice as fast.

You do not need the thread to establish any of this, which is the genuinely unusual part. Every model repo ships a THIRD_PARTY_NOTICES.md that names its parent, its license and the exact modifications. Voz's version is about as blunt as a legal document gets:

Weight values are otherwise unchanged from the base model.

What Desert Ant did do to Parakeet: convert it out of NeMo into compiled Core ML, compress it with 6-bit palettization, and restructure it into separate encoder, frontend and decoder graphs that their Swift runtime dispatches. That drops it to 467MB on disk against 1.6GB for Whisper large-v3-turbo at fp16, and it is what buys the 4.7x speedup over whisper.cpp on the same podcast audio.

Think of the weights as the recipe and the runtime as the kitchen. Desert Ant did not write a new recipe for Voz. They built a kitchen that cooks the existing one about five times faster, on a phone, for free. That is worth money. It is just not the same claim as "we trained the models ourselves," which is what the launch post says two paragraphs earlier.

The models nobody talked about are the good ones

Tongue identifies which of 84 languages a piece of text is in, from three words, out of a 2MB file. Its notices file opens with a sentence the others cannot: trained from scratch, derives from no third-party model, corpus assembled from CC0 and CC BY sources with a provenance manifest recording every file kept and dropped. It scores 0.933 where a 293MB detector scores 0.887.

Redact is the better story. It masks names, addresses, cards, IBANs and national IDs across 27 languages before text leaves the device, from 23M parameters, using Microsoft's Multilingual-MiniLM truncated to six layers and distilled with GLiNER-PII and DeepSeek-V3.2-Exp as teachers. On Desert Ant's harness it catches 88.8% of personal data at 99.6% precision from a 12MB Apple build. GLiNER-PII, its own teacher, catches 91.1% at 90.4% precision and weighs 2.3GB.

The numbers that hold up, and the ones with an asterisk

Credit where it is due: Desert Ant publishes its own asterisks, which is rarer than it should be.

• Voz accuracy is slightly behind Whisper, and the page says so. Across six Open ASR Leaderboard sets it averages 7.40% word error rate against Whisper large-v3-turbo's 7.00%, using Voz's own text normalizer. It wins exactly one row, AMI at 11.84% against 13.87%, which happens to be the meeting-room audio most products actually see.
• Clear has no published quality score. The launch post says Desert Ant replaced Dolby with it. The Clear model page says, in its own words, that they have not published an audio-quality score against the cloud tools. Both statements are on the same website.
• The energy comparison is an estimate stacked on an assumption. Clipping 100,000 thirty-minute videos costs 0.3 kWh with Clips against 140 kWh with Claude Sonnet, per the about page. The Sonnet figure is estimated from published MLPerf runs, and the Clips figure assumes an iPhone 17 Pro draws 3W. The direction is obviously right. The 470x is not a measurement.
• The iPhone timings are internal. Their words again: the phone numbers and the whisper.cpp comparison are their own runs and are not on the model card yet.

Read the license before you read the benchmarks

This is where the launch stops being a weights story and becomes a business-model story. The Desert Ant Labs Source-Available License 1.0, dated July 3, says out loud that it is source-available and not open source. The terms that matter if you are shipping:

• Free below 100,000 monthly active devices, counted per model per platform. The same model on iOS and Android gets two separate allowances. Above that you email licensing. Research and teaching are exempt entirely.
• You may not train a competitor on it, including on the outputs, the logs, the embeddings or the evaluations. That covenant survives termination.
• You may not redistribute the models standalone. Embed and ship them in your app, which is the whole point of the license, but you cannot repackage them as an SDK or a hosted service.
• You owe visible attribution, a "Powered by Desert Ant Labs" line somewhere your users can see.
• The SDK phones home. Section 7 describes minimized telemetry using an SDK-generated identifier that counts monthly active devices and carries no user content. Section 6 makes tampering with it a breach.

So the on-device lab whose pitch is that nothing leaves the device counts your devices from a server. That is not hypocrisy, it is just how you meter a free tier you cannot see, and Desert Ant documents it in more detail than most cloud vendors document their logging. Worth knowing before your privacy page goes to legal.

The underlying argument is still correct

Desert Ant leans on Small Language Models are the Future of Agentic AI, the NVIDIA position paper, for the claim that 40 to 70% of an agent's model calls could go to a small specialized model. That checks out: the paper's three case studies estimate about 60% for MetaGPT, 40% for Open Operator and 70% for Cradle. If you are paying per token to tag a topic, catch a filler word or spot a language, you are buying a philosophy degree to read a thermometer.

The practical takeaway is not "avoid Desert Ant." Several of these models are the fastest way to ship a feature today, and the packaging work is the part almost nobody does well. It is that the notices file is now the first file you open in any model repo. It tells you whether you are buying weights, a runtime, or a wrapper you could build yourself in a weekend with a frontier model and a spare Mac, which one commenter in the thread pointed out is a very hill-climbable task these days.

Key Takeaways

• Desert Ant Labs shipped 18 on-device models on September 8 with Swift, Kotlin and JavaScript SDKs, a CLI, and weights on Hugging Face, free below 100,000 monthly active devices per model per platform.
• Voz is NVIDIA's Parakeet TDT 0.6B v3 with the weight values unchanged, converted to Core ML and 6-bit palettized. Clear fine-tunes DeepFilterNet 3, Ear is a whisper-tiny subgraph, Clips is an XLM-RoBERTa trunk with new heads. Every one of these is disclosed in the repo's own notices file.
• The runtime is the actual product. 319x realtime transcription on an M3 Ultra against 78x for Apple's SpeechAnalyzer, and 467MB on disk against Whisper turbo's 1.6GB, come from compilation and packaging rather than better weights.
• The strongest original work went unmentioned: Tongue at 2MB trained from scratch, and Redact at 23M parameters landing 88.8% PII recall at 99.6% precision against a 2.3GB teacher's 91.1% at 90.4%.
• It is source-available, not open source. No competing-model training, no standalone redistribution, mandatory visible attribution, and an SDK that reports a device count back to Desert Ant.
• Open the notices file first. It is the cheapest way to tell a new model from a new build of an old one.

Sources: Desert Ant Labs launch post, Voz third-party notices, Voz model page, Clear third-party notices, Clear model page, Ear third-party notices, Clips third-party notices, Redact third-party notices, Redact model page, Tongue third-party notices, Desert Ant Labs Source-Available License 1.0, Desert Ant Labs about page, desert-ant-core SDK, NVIDIA Parakeet TDT 0.6B v3, Small Language Models are the Future of Agentic AI (arXiv 2506.02153), Hacker News discussion