TL;DR:

  • MediaTek’s Genio SoC family brings on-device AI — 6 to 10 TOPS depending on variant — to retail POS, inventory, robotics, and industrial systems
  • The Genio 360 series supports generative AI models up to 2 billion parameters without a cloud connection
  • Key use cases: frictionless checkout, real-time shelf monitoring, voice-driven POS, autonomous industrial inspection
  • The platform runs Linux and Android, with a unified software architecture across the product line

Most AI in deployed IoT systems today is either very limited — inference of simple classification models — or dependent on a cloud round-trip for anything substantial. MediaTek’s Genio platform is positioned at the seam between these: enough on-device AI compute to run small generative models and sophisticated vision tasks, in SoCs designed for the power envelope and industrial grade requirements of embedded IoT hardware.

MediaTek highlighted the Genio platform at NRF 2026 for retail applications and has since expanded the product line to cover robotics, drones, and industrial IoT. Here is what the platform looks like across its current variants.

Genio 360 and 360P: Entry-Level Generative AI at the Edge

The Genio 360 and 360P are the most accessible entry points in the current Genio lineup. The standard 360 delivers 6 TOPS of system-level AI performance; the 360P steps up to 8.5 TOPS. Both support generative AI models with up to two billion parameters on-device — enough for capable instruction-following, classification, anomaly detection, and lightweight vision-language tasks without sending data to a server.

Display support covers single 4K60 or dual FHD60 configurations, with advanced video encode/decode included. Memory supports up to 8GB of LP4X-3733. The target applications are high-end digital signage, interactive kiosks, and smart retail terminals where the device needs to run AI-driven interactions locally with no latency to a cloud inference endpoint.

For retail specifically, the Genio 360 family enables self-checkout kiosks that use on-device vision AI to verify items, contactless payment terminals that process facial or voice biometrics locally, and digital signage that adapts content based on on-device audience detection without transmitting camera feeds.

Genio 520 and 720: Workhorses for Connected Retail

The Genio 520 and 720 series step up to a 6nm octa-core processor pairing two ARM Cortex-A78 cores at up to 2.2 GHz with six ARM Cortex-A55 cores at up to 2.0 GHz. The integrated NPU 850 delivers 10 TOPS of AI performance — a meaningful increase for more demanding vision inference and faster model response times.

Fibocom’s AI ECR (Edge Computing Reference) solution, demonstrated at MWC 2026, is built on the Genio 520/720 platform. The hardware handles traditional POS workloads alongside AI use cases — inventory scanning, real-time stock counting via overhead camera, voice order processing — on a single SoC without requiring separate AI co-processors.

The 10 TOPS headroom allows more complex vision models. Object detection models capable of identifying out-of-stock shelf positions in real time, tracking customer flow patterns through a store, and performing product recognition at checkout without barcodes all sit within this compute envelope.

Genio 700: Power-Efficient Performance for Demanding Applications

The Genio 700 targets applications that need sustained AI compute with strict power constraints — robotics, drone vision systems, and industrial equipment where thermal dissipation is limited. The 6nm-class SoC combines an octa-core CPU with two Cortex-A78 processors at 2.2 GHz, an Arm Mali-G57 GPU, and an AI processor delivering 4.0 TOPS.

The 4.0 TOPS figure is lower than the 360 and 520/720 on paper, but Genio 700’s differentiation is in power efficiency and industrial grade certification — important for equipment running continuously in factory environments where thermal management and reliability matter more than peak performance.

Software Ecosystem

One of the engineering advantages of the Genio platform is its unified software architecture across the product line. MediaTek provides a consistent SDK that abstracts the hardware differences between Genio variants, which means software developed for a Genio 360 application can be adapted to Genio 700 without rebuilding the AI inference stack.

The platform runs both Linux and Android, with MediaTek maintaining Yocto-based Linux BSPs for industrial applications and Android BSPs for retail and consumer deployments. OEM and ODM partners have developed reference boards for the major Genio variants, accelerating time to deployment for custom hardware designs.

For AI model deployment, the Genio SDK includes MediaTek NeuroPilot, a unified AI inference framework that handles model conversion, optimization, and hardware-accelerated execution across the NPU, CPU, and GPU. Models from TensorFlow, PyTorch, and ONNX are supported through the NeuroPilot pipeline.

Real-World Use Cases Already Deployed

MediaTek’s retail demonstrations at NRF 2026 focused on four use cases that are in active deployment on Genio-based hardware:

Shelf organization and real-time inventory tracking: Overhead cameras running vision models on-device detect shelf gaps, misplaced products, and low-stock conditions in real time, with alerts generated locally without cloud latency.

Frictionless payments: On-device facial recognition and biometric processing enable payment authorization without card insertion or NFC, processed at the terminal without biometric data leaving the device.

Voice-driven POS: On-device speech recognition handles order entry in noisy retail environments, again without the round-trip latency of cloud speech APIs.

AI-driven customer interaction: Interactive kiosks using on-device LLMs (sub-2B parameter models on Genio 360, larger on 520/720) handle product queries, recommendations, and wayfinding without connectivity requirements.

Who Should Be Paying Attention

The Genio platform is most relevant for three audiences. IoT product teams evaluating SoCs for next-generation intelligent devices will find Genio’s compute-per-watt ratio competitive with alternatives, particularly for applications that need both display output and AI inference in the same device.

System integrators building smart retail or industrial automation solutions face increasing customer expectations around on-device AI — data residency requirements, latency constraints, and offline operation are all pushing against cloud-dependent architectures. Genio provides a credible on-device path with sufficient headroom for 2026 AI model sizes.

Finally, enterprises deploying AI at scale across distributed locations — hundreds or thousands of devices — find the economics of on-device inference compelling compared to per-query cloud costs. Genio’s cost structure as a commodity SoC makes the per-device AI compute cost predictable and independent of usage volume.

Sources