Developers looking to embed artificial intelligence directly into applications without token costs or inference latency now have access to a new suite of edge models. European AI research lab Desert Ant Labs has released 18 specialized on-device models designed to process voice, image, and text locally on consumer hardware. The release comprises 12 stable versions and 6 beta versions, each tailored for dedicated tasks. Integration is supported via software development kits for Swift, Kotlin, and JavaScript, with free tier access provided for up to 100,000 monthly active devices.
Local Audio Processing and Performance Benchmarks
The audio enhancement model, Clear, weighs in at 9 megabytes and transforms 5 minutes of laptop microphone recordings into studio-quality audio in 1 second. The model handles audio cleanup, mastering, and re-encoding entirely on the device. In benchmarks evaluating 5-minute audio processing tasks, the iPhone 16 Pro achieved a processing speed of 302 times real-time, while a MacBook Pro equipped with the M5 chip reached 345 times real-time performance.
For transcription tasks, the flagship model Voz transcribes 10 minutes of audio on an iPhone in 2 seconds, proving 4.7 times faster than Whisper. Voz also supplies exact start and end timestamps for every word. On an M3 Ultra system processing 30 consecutive minutes of audio, the model recorded a speedup of 319 times real-time. By comparison, Apple SpeechAnalyzer reached 78 times real-time and Whisper large-v3-turbo hit 50 times. Running on an iPhone 17 Pro environment, the system achieved a processing speed of 298 times real-time.
Privacy Redaction and Real-World Application Integration
Text security and privacy protection are handled by Redact, a 12-megabyte model built to detect names, addresses, and credit card numbers across 27 languages in real time. The model independently masks 88.8% of personal information in input text, completely preventing sensitive data from ever reaching a remote server. For comparison, GLiNER-PII spans 2.3 gigabytes with a 91.1% detection rate, Rampart measures 14.7 megabytes with a 61.4% score, and the OpenAI filter occupies 3 gigabytes while delivering 60.2% performance.
The video application Detail faced escalating infrastructure costs alongside a lack of local model options as its user base expanded. The engineering team spent months searching Hugging Face for models capable of filler-word detection or recording cleanups that could be dropped directly into production. Finding no suitable alternatives, they built their own models, replacing Dolby audio enhancement with Clear. They also swapped out Claude Sonnet for Clips, their own 284-megabyte model, to boost both speed and energy efficiency. Detail 6 will ship alongside iOS 27, completely transitioning cloud API dependencies to on-device execution models.
Analysis from NVIDIA researchers indicates that across three distinct agent systems, 40 to 70 percent of large model API calls can be offloaded to small specialized models. Rather than querying massive cloud APIs for repetitive tasks, running task-optimized small models locally eliminates server round-trip latency and communication overhead. With over a billion smartphones, tablets, and laptops shipping globally, the aggregate compute capacity sitting inside user-owned hardware now surpasses traditional AI data center throughput, driving this architectural shift toward local edge execution.
