Desert Ant Labs Releases 18 Ultra-Fast, Small AI Models for On-Device Execution

Europe-based AI research lab Desert Ant Labs has released 18 small specialized models (12 Stable, 6 Beta) that run directly on-device to reduce reliance on cloud APIs. These models are optimized for audio, vision, and text processing and are free for up to 100,000 monthly active devices.

Key Model Performance and Features

Each model focuses on specific tasks, offering overwhelming speed and efficiency compared to existing cloud models.

• Voz (Audio Transcription): Processes 10 minutes of audio on an iPhone in 2 seconds, which is 4.7x faster than Whisper. On the M3 Ultra, the real-time factor is 319x, significantly outperforming Apple SpeechAnalyzer (78x) and Whisper large-v3-turbo (50x).
• Clear (Audio Enhancement): A lightweight 9MB model that converts 5 minutes of laptop recording to studio quality in 1 second. On the MacBook Pro (M5), the real-time factor reaches 345x.
• Redact (PII Masking): Masks names, addresses, card numbers, etc., in real-time across 27 languages. The 12MB model achieves a personal data detection accuracy of 88.8%, showing competitive performance against the larger GLiNER-PII (91.1%).
• Tongue (Language Identification): A 2MB model that identifies 84 languages with just 3 words, achieving an accuracy of 0.933.
• Clips (Video Clipping): Generates 12 clips from a 10-minute video in 5 seconds. It is 10x faster than Claude Sonnet and uses 470x less energy.

Strategy and Deployment

Desert Ant Labs prioritized the development of always-on specialized models based on the 'cerebellum' concept to address infrastructure cost increases and latency issues in existing cloud services. In the future, they plan to introduce a 'cortex' structure that hierarchically routes from small local models to large cloud models.

Developers can integrate these models via a single SDK for Swift, Kotlin, and JavaScript, and try them out via the CLI for Mac and Hugging Face. Additionally, their own video editing app, Detail 6, will launch alongside iOS 27 and is expected to replace all cloud APIs with their own on-device models.