欧洲前沿AI实验室Desert Ant Labs宣布推出18个本地运行的小型专用AI模型，可在手机、平板等设备上离线执行任务，无需调用云API1。这些模型涵盖音频转录、音质增强和隐私脱敏等多项功能，每月100万活跃设备以下免费使用，不收取token费用1。
产品方面，该实验室推出的12个稳定版和6个测试版模型展现出显著的性能优势1。其中Voz音频转录模型在iPhone上处理10分钟音频仅需2秒，速度比Whisper快4.7倍；Clear音质增强模型体积仅9MB，将5分钟录音升级至工作室质量耗时1秒；Redact隐私脱敏模型支持27种语言实时处理；Clips视频转换模型在5秒内将10分钟视频转为12个短视频，相比Claude Sonnet快10倍，能耗低470倍1。为方便开发者集成，Desert Ant Labs提供了Swift、Kotlin和JavaScript SDK1。
根据NVIDIA研究估计，代理系统中40-70%的大模型调用可被小型专用模型替代1。
Desert Ant Labs, a European AI research laboratory, has announced the release of 18 specialized small-scale models designed to run locally on devices such as smartphones and tablets without requiring cloud API calls 1. The models, which include 12 stable versions and 6 experimental versions, cover tasks spanning audio transcription, audio enhancement, and privacy redaction 1.
The company's Voz audio transcription model achieves performance 4.7 times faster than Whisper, processing a 10-minute audio file on an iPhone in just 2 seconds 1. Another offering, the Clear model, requires only 9 megabytes of storage and upgrades a 5-minute recording to studio quality in 1 second 1. The Redact model supports real-time de-identification across 27 languages 1. Additionally, the Clips model, at 284 megabytes, can convert a 10-minute video into 12 short videos in 5 seconds—10 times faster than Claude Sonnet while using 470 times less energy 1.
Desert Ant Labs offers its models free of charge for devices with fewer than 1 million active monthly users, with no token-based billing 1. The company provides Swift, Kotlin, and JavaScript SDKs to enable developer integration 1. According to NVIDIA research cited by the company, 40 to 70 percent of large model calls in agentic systems could be replaced by specialized smaller models 1.
评论
还没有评论，欢迎留下第一条。