Desert Ant Labs deploys its models directly into applications

Desert Ant Labs launches 18 specialized models to process audio, text, and images on-device, without sending content to the cloud.

Transcribing ten minutes of audio in two seconds, cleaning up a five-minute recording in one second, or masking personal data as it is typed: Desert Ant Labs says it can perform each of these operations directly on a phone or computer, with no per-request billing and without sending the content to a remote server.

Introduced on September 8, the company describes itself as a European frontier AI lab. Its first catalog includes 18 models dedicated to specific audio, text, and image tasks. Twelve are currently available, while six remain in beta, with no public SDK at launch.

This distinction qualifies the claim that all 18 models are already “live.” Align, Clear, Clips, Ear, Emo, Gist, Redact, Shapes, Title, Tongue, Uhm, and Voz can be integrated today. Eye, Face, Moderator, Schemer, Toxic, and Who appear in the catalog, but the documentation still lists them as coming soon and invites developers to request early access.

For now, the company is not trying to build a general-purpose assistant capable of handling any request. Each model has a narrowly defined role: recognizing speech, enhancing a voice recording, finding the best moments in a transcript, identifying a language, masking personal information, or suggesting a title.

This specialization makes it possible to reduce file sizes and processing requirements considerably. Tongue, which identifies languages, is just 2MB. Clear takes up 9MB, while Redact occupies 12MB on Apple devices. At the larger end of the catalog, Voz and Clips weigh in at 467MB and 284MB respectively, but they remain compact enough to be bundled with an app or downloaded on demand.

Desert Ant Labs provides SDKs for Swift, Kotlin, and JavaScript. That coverage applies to the product line as a whole, but not necessarily to every model. Voz, for example, currently works through Swift on Apple devices equipped with the company’s own chips. Clear and Redact have broader implementations covering Android and the web as well.

The company grew out of the team behind Detail, a video creation app developed over the past five years. Some Detail features previously relied on remote services to enhance sound, transcribe recordings, or select clips. As the app gained users, the cost of those calls increased.

The team says it began training and adapting its own models to replace those services. Clear reportedly replaced a Dolby tool for audio enhancement, Voz accelerated local transcription, and Clips took over from Claude Sonnet in the automatic editing system. Detail 6, scheduled to launch alongside iOS 27, is expected to eliminate the remaining calls to remote AI services used by the app.

That experience explains the company’s strong focus on practical products. Desert Ant Labs is not simply distributing files for developers to download, but combining models, runtimes, and programming interfaces tailored to each platform. Clear and Voz use the Neural Engine in Apple devices, while the web version of Clear runs the same weights through WebAssembly.

Voz is one of the collection’s main technical showcases. The model can transcribe speech in 25 European languages, with start and end timestamps for every word. On an iPhone 17 Pro, Desert Ant Labs reports two seconds for ten minutes of audio and six seconds for thirty minutes. An iPhone 15 Pro reportedly completes the same thirty-minute transcription in seven seconds.

The model is based on NVIDIA’s Parakeet TDT 0.6B v3, released under the CC BY 4.0 license. Desert Ant Labs says it did not modify the weights. Its contribution lies in the conversion, compression, and runtime designed for the Neural Engine.

The comparison with Whisper requires some caution. The reported 4.7-fold speed advantage was measured by the company on an M3 Ultra Mac, against Whisper large-v3-turbo running through whisper.cpp. It should not be treated as a guaranteed advantage across every device, implementation, or type of recording.

Voz is not consistently more accurate either. Across six public English speech-recognition datasets, its average word error rate is 7.40%, compared with 7% for Whisper large-v3-turbo. It performs better on meetings from the AMI dataset, scoring 11.84% versus 13.87%, but remains slightly behind Whisper on several other datasets.

Results vary more widely across languages. In the company’s published measurements, the word error rate ranges from 3.31% for Italian to 39.46% for Greek. Voz also does not automatically identify the spoken language. A recording in an unsupported language may produce a transcript that appears plausible but is incorrect. Desert Ant Labs recommends pairing it with Ear, its spoken-language detection model.

Voz’s 467MB package must be downloaded and stored on the device. After installation, an initial Core ML preparation step can take about twenty seconds. Subsequent loads reportedly fall to around 0.2 seconds. These details illustrate one of the trade-offs of local processing: server latency and cloud fees disappear, but the app must manage storage, hardware compatibility, and model initialization.

Clear applies the same approach to voice enhancement. Its 9MB file removes noise, adjusts loudness, and re-encodes a recording at 48kHz. Two versions are available. Clear Studio attempts to remove background noise almost completely, while Clear Natural preserves more room tone, breathing, and vocal texture.

Desert Ant Labs reports processing speeds of 302 times real time on an iPhone 16 Pro and 345 times real time on an M5 MacBook Pro. A five-minute recording could therefore be processed in about one second. These figures, however, come from internal measurements conducted with the shipping version of the software. The company has not yet published a perceptual quality score comparing the output with remote services, and it notes that Android and web performance has not been measured under the same conditions.

Redact targets another benefit of local processing: preventing personal information from ever reaching a server. The model