AIBOX-9075 Brings Large AI Models to the Edge
Originally published by the Firefly Open Source Team | August 27, 2026 | Guangdong, China
As AI moves from the cloud to real-world devices, computing power, models, and development tools are all essential. Built on high-performance hardware, supported by a rich model ecosystem, and equipped with the complete AidLux toolchain, AIBOX-9075 lowers the barrier to development and helps teams deploy large AI models and vision AI applications directly at the edge.
Powerful Computing for Demanding AI Workloads
AIBOX-9075 is powered by the Qualcomm Dragonwing IQ9-series flagship industrial SoC, the IQ-9075. It integrates an eight-core, all-big-core Kryo Gen6 CPU and delivers up to 200 TOPS of peak INT8 AI performance for compute-intensive edge AI workloads and concurrent tasks. With CPU, GPU, and dual Hexagon NPU acceleration, it supports private, on-device deployment of large models and generative AI. Large-model inference performance reaches up to 12 TPS with a 13B-parameter model, meeting the needs of edge-based Q&A, knowledge bases, intelligent assistants, and similar applications.
|
Item |
Specification |
|
Product |
AIBOX-9075 industrial-grade edge computing box |
|
SoC |
Qualcomm Dragonwing IQ-9075 |
|
CPU |
Eight-core Kryo Gen6, up to 2.36 GHz |
|
GPU |
Adreno 663 |
|
NPU |
Dual Hexagon NPU |
|
Peak AI performance |
200 TOPS INT8 |
|
General-purpose performance |
230 KDMIPS |
|
Process |
5 nm LPE |
|
Large-model inference |
Up to 12 TPS with a 13B-parameter model |
Broad Model Support Across AI Applications
Through the AidLux Model Farm, developers can access a large collection of optimized models and sample code across the following categories:
· Computer vision: YOLO series, SAM2, OCR, pose estimation, image segmentation, depth estimation, and more.
· Natural language processing: Qwen, Llama, DeepSeek, Gemma, Phi, and other large language models.
· Multimodal AI: Qwen2.5-VL, LLaVA, InternVL, CLIP, and more.
· Speech and audio: Whisper, WeNet, SenseVoice, MeloTTS, and more.
The platform supports multiple precision formats, including FP16, INT8, W8A16, and W4A16, allowing developers to balance accuracy, speed, and resource usage. Model Farm also provides one-stop access to model search, performance references, downloads, and sample code, significantly reducing the effort required for model selection and adaptation.
Faster Deployment, Easier Development
1. Familiar Linux Development Environment
AIBOX-9075 runs Ubuntu 24.04 and supports both native compilation and ARM64 cross-compilation. Developers can use familiar tools such as GCC, CMake, Python, and OpenCV, making it suitable for rapid validation, on-site debugging, and production development.
2. Complete AidLux Software Toolchain
The out-of-the-box AI development and deployment toolchain integrates ten mainstream AI frameworks. A unified AidLite API automatically identifies models and matches them with the appropriate framework, enabling AI scenarios with only a few lines of code. A drag-and-drop Wizard tool supports no-code development, while Python alone can be used to build and deploy AI applications. Together, these capabilities lower the development barrier and shorten time to deployment.
AIMO: Model conversion and optimization.
Model Farm / MMS: Model search, download, and management.
AidLite: Unified access to CPU, GPU, and NPU resources for common AI model inference.
AidGen: On-device inference for large language models and multimodal models.
AidGenSE: An OpenAI API-compatible HTTP service for easy integration with existing applications.
AidVoice: Speech recognition and speech synthesis.
AidStream: GStreamer-based processing for real-time video streams.
Intelligent Scene Analysis Platform Demo
The platform supports multiple types of video sources and uses vision-language models to understand on-screen content. Users can define preset conditions for different recognition scenarios in natural language. It is suitable for home security, production quality control, public safety, and other video-monitoring and recognition applications. The source article includes a real-world demonstration video.
Hardware and Industrial-Grade Adaptability
The system integrates GMSL2, CAN-FD, RS485, and optically isolated DI/DO interfaces. Its industrial wide-temperature design supports stable operation from -40°C to 85°C, enabling deployment in factories, vehicles, robots, outdoor sites, and other demanding environments.
Interface and Connector Reference
|
Item |
Specification |
|
Debug and storage |
USB Type-C debug port; SIM card slot; flash button; status indicator; power button |
|
Power |
12 V DC input, 5.5 x 2.1 mm |
|
USB |
2 x USB 3.0 ports, 1 A current limit |
|
Ethernet |
2 x 2.5GbE Ethernet ports |
|
Vision capture |
8 x GMSL2 channels using 4-pin Mini FAKRA connectors |
|
Industrial I/O |
2 x CAN-FD; 2 x RS485; optically isolated DI/DO |
|
Terminal block |
36-pin Phoenix terminal block, 3.5 mm pitch |
|
Display |
HDMI 2.0, up to 4K at 60 fps |
|
Wireless antennas |
4 x 5G antenna connectors; 2 x Wi-Fi antenna connectors |
GMSL2 is an automotive-grade serial-link technology designed for high-bandwidth video transmission. It supports up to eight cameras and long-distance transmission, allowing cameras to be deployed far from the central computing unit and simplifying installation in large equipment or vehicles. Low latency and strong noise immunity help maintain reliable data transmission in industrial environments and other locations with complex electromagnetic conditions.
Typical Application Scenarios
Industrial inspection: Product defect detection, OCR, PPE compliance detection, and equipment-status recognition.
Intelligent monitoring: Analysis of people, vehicles, intrusions, abnormal behavior, and multiple video streams.
Robotics: Visual perception, scene understanding, voice interaction, and task control.
On-device large-model assistant: Local knowledge bases, equipment maintenance Q&A, and on-site work guidance.
Voice interaction: Voice control, speech transcription, and local voice assistants.
An Edge AI Platform Built for Real-World Deployment
AIBOX-9075 is more than a high-performance computing device; it is an integrated hardware and software platform built to bring edge AI into production. Its 200 TOPS peak performance supports complex models. Its broad model ecosystem spans vision, language, multimodal AI, and speech. The AidLux toolchain reduces the development barrier with a clear workflow and comprehensive tools from model selection through application integration.
For developers and enterprises, this means a faster end-to-end path from model validation and application development to on-site deployment, turning AI capabilities into production-ready products.
High Performance · Broad Model Support · Complete Toolchain
Empowering Real-World Edge AI Deployment
References and Contact
· AIBOX-9075 product introduction
· AIBOX-9075 official documentation
· AidLux software guide
· AidLux Model Farm
Contact Firefly for additional industry application solutions:
Email: sales@t-firefly.com
Website: www.t-firefly.com
Follow the Firefly Open Source Team on WeChat Official Account, WeChat Channels, and bilibili. Official store: Firefly Direct Store.
Previous article: Full-Stack AI in One Box: RK1828 Edge Multimodal Inference Platform
