MeiG Smart Brings 7B-Class Generative AI to the Edge with Dragonwing IQ-9075
Why MeiG Smart built on Dragonwing IQ-9075
MeiG Smart needed a platform that could support embedded AI without forcing customers into a high-power, GPU-centric system design. The goal was not just model throughput. The platform also had to support camera input, display output, networking, storage, industrial I/O, thermal constraints, and customer application logic at the same time.
The product requirements included:
- Run LLMs, vision-language models, and vision AI locally
- Leave CPU headroom for the customer application
- Keep power and thermals under control
- Support camera, display, networking, storage, and industrial I/O
- Support automotive and industrial environments
- Provide memory capacity for larger AI models and multi-application workloads
Dragonwing IQ-9075 provided MeiG Smart the compute architecture for this approach: CPU, GPU, NPU, memory, and I/O integrated into a platform that can run AI workloads while still supporting the rest of the embedded system.
Dragonwing IQ-9075 presented MeiG Smart with a more balanced architecture. The Qualcomm® Hexagon™ NPU can carry AI inference while the CPU remains available for system workloads. Memory capacity was especially important for local generative AI. Dragonwing IQ-9075 supports up to 36 GB LPDDR5X depending on configuration, which affects model size, context length, vision-language workloads, and the ability to run inference alongside the customer application.
The proof: A Qwen2.5 7B generative model running on the platform
The most demanding workload MeiG Smart put on the platform is also the clearest evidence. On the SNM983, running on the Hexagon NPU with W4A16 quantization, three different 7-billion-parameter models came up successfully and ran fully on-device, with no cloud in the loop.
Overall, the CPU column is the key metric to focus on. The NPU carries the language model, while the CPU remains around one-third utilized. That means a customer can run a 7B assistant and still have compute, memory, and thermal budget available for the robot, camera pipeline, user interface, networking stack, and application logic on the same chip.
A GPU-centric design that uses most of the SoC to reach similar throughput can leave less room for the product workload around the model. MeiG Smart's result shows why NPU-first inference matters for embedded systems: the model runs locally, but the system remains usable.
MeiG Smart expects more, as the above results are early and conservative numbers. With Hexagon optimizations such as multi-head-attention splitting, NCHW-to-NHWC layout conversion, operator substitution, and fuller use of dual-HTP capacity, MeiG Smart expects higher prefill and decode throughput from the same hardware.
MeiG Smart Dragonwing IQ9-based Product Portfolio: SOM and AI Box
SNM983 SOM
The SNM983 is designed to provide the compute foundation for products that need on-device AI, camera input, display output, high-speed storage, networking, USB, PCIe, UART, SPI, I2C, GPIO, and other embedded interfaces. The SOM approach aims to give engineering teams control over the carrier board, enclosure, I/O, thermal design, and application-specific integration while avoiding a full compute-module design from the ground up. It is the design-in product. It is built for companies that want to integrate Dragonwing IQ-9075 into their own system architecture.
MT200 AI Box
The MT200 integrates the SNM983 SOM into a compact industrial edge AI computing platform. It is designed for teams that want a complete AI system for evaluation, development, pilots, or production deployment. The AI Box is positioned for Linux and Ubuntu-based development, remote debugging, model deployment, FOTA updates, package management, and low-code workflows. It was developed to give customers a way to start with a working edge AI platform before committing to a custom design.
Target Use Cases
1. Edge AI Box
On-device model deployment for LLMs, vision-language models, and security-focused AI assistants. Supports local inference for Qwen, Llama, and similar models without relying on the cloud.
2. Industrial Detection
Multi-camera inspection with real-time AI detection. Common applications include defect detection, production-line monitoring, OCR, dimensional inspection, material sorting, and behavior detection.
3. Robotics
Multi-camera perception with low-power autonomous navigation. Supports robot vision, visual SLAM, obstacle detection, path planning, object recognition, voice interaction, and 16 GMSL2 cameras depending on system design.
4. IVI and Automotive AI
On-device AI inference and prediction for in-vehicle displays. A Tier-1 is already working with the platform for an IVI program at a major automotive OEM.
5. Mini PC
Compact local AI compute for Mini PC designs. A Middle East project is currently under evaluation, with use cases including local LLMs, private AI assistants, and edge productivity workloads.
Key Takeaway
MeiG Smart is not only building another edge AI module. It is building a path from embedded AI design-in to deployable AI systems.
The SNM983 provides OEMs a Dragonwing IQ-9075 SOM for product integration. The MT200 provides developers and system integrators a complete AI Box using that same compute foundation. The 7B model results show the platform can already run serious generative AI locally while leaving CPU and memory headroom for the rest of the product.
The main goal: On-device AI that can fit into automotive, industrial, robotics, vision, Mini PC, and enterprise systems without turning the entire system into a model accelerator and nothing else.
