Boca Raton, Fla. – – September 11, 2026 -- Telit Cinterion unveiled an edge AI software development kit that runs machine learning inference directly on the application core of its cellular modules, eliminating the need for a separate AI accelerator chip. The IoT connectivity provider said proof-of-concept testing on image classification and object detection tasks kept CPU consumption at or below 17%, a threshold that avoids thermal throttling of 4G and 5G radio performance.
SDK embeds Google's LiteRT runtime into module firmware set for Q4 2026 release
The kit incorporates LiteRT, the runtime formerly known as TensorFlow Lite, into the Linux-based firmware of upcoming Telit Cinterion AI-enabled module variants spanning 4G, 5G RedCap and high-performance 5G. The company plans to ship the SDK in the fourth quarter of 2026.
Standard .tflite format lets developers skip proprietary rebuilds
Because LiteRT runs the industry-standard .tflite file format, developers can train and optimize models in existing tools and load them onto a Telit Cinterion module without a vendor-specific rebuild. A model already running on a PC or a Raspberry Pi transfers directly to the module, according to the company.
Compact runtime brings on-device inference to low-cost, connectivity-only hardware
The SDK's small footprint allows AI processing on modules previously too constrained to host larger AI software stacks. Included sample applications handle the full inference pipeline -- acquiring sensor data, preprocessing it, running the model and returning predictions to the industrial application -- while system integrators retain control over the final model and deployment architecture.
Target use cases include predictive maintenance and smart metering
Telit Cinterion identified three customer application areas: predictive maintenance using vibration or audio analysis to flag equipment anomalies in motors, pumps and bearings before failure; acoustic monitoring to detect alarms or breaking glass at remote industrial sites; and smart metering that reads existing analog meters via connected cameras without replacing installed hardware.
"Industrial IoT teams should not have to redesign their entire device architecture to add practical AI capabilities," said Marco Argenton, senior vice president of product management at Telit Cinterion. Vishal Batra, the company's vice president of software engineering, said the engineering challenge lies in delivering secure and reliable intelligence within the constraints of a connected embedded device, not merely running AI at the edge.